investment dispersion and the business cycle - uni … · investment dispersion and the business...
TRANSCRIPT
Investment Dispersion and the Business Cycle
Rudiger Bachmanna,∗ Christian Bayerb,†‡
aRWTH Aachen University, Templergraben 64, Rm. 513, 52062 Aachen, Germany.
bUniversity of Bonn, Adenauerallee 24-42, 53113 Bonn, Germany.
June 3, 2013
Abstract
The cross-sectional dispersion of firm-level investment rates is procyclical. Thismakes investment rates different from productivity, output and employment growth,which have countercyclical dispersions. A calibrated heterogeneous-firm business cyclemodel with nonconvex capital adjustment costs and countercyclical dispersion of firm-level productivity shocks replicates these facts and produces a correlation betweeninvestment dispersion and aggregate output of 0.53, close to 0.45 in the data. We findthat small shocks to the dispersion of productivity, which in the model constitutes firmrisk, suffice to generate the mildly procyclical investment dispersion in the data but donot produce serious business cycles.
JEL Codes: E20, E22, E30, E32.Keywords: Ss model, RBC model, cross-sectional firm dynamics, lumpy investment,aggregate shocks, idiosyncratic shocks, risk shocks, heterogeneous firms.
∗Corresponding author. Phone: +49-241-8096203. Fax: +49-241-8092649. E-mail:[email protected]. Also: NBER (United States), CESifo (Germany), and ifo (Germany).†E-mail: [email protected].‡We thank Dirk Kruger, Giuseppe Moscarini, Matthew Shapiro as well as four anonymous referees for their
comments. We are grateful to seminar/meeting participants at Amsterdam, the ASSA (San Francisco), Bonn,Duke, ECARES (Brussels), ESSIM 2009, Georgetown, Groningen, Johns Hopkins, Innsbruck, Michigan-AnnArbor, Michigan State, the Minneapolis Fed, Notre Dame, Regensburg, the SED (Istanbul), Wisconsin-Madison and Zurich for their comments. We thank the staff of the Research Department of DeutscheBundesbank for their assistance. Special thanks go to Timm Koerting for excellent research assistance. Thispaper formerly circulated under the current title as NBER-WP 17063 and under: “The Cross-section ofFirms over the Business Cycle: New Facts and a DSGE Exploration” as CESifo-WP 2810. Any remainingerrors are our own.
This paper establishes a novel business cycle fact: the cross-sectional dispersion of firm-
level investment rates is robustly and significantly positively correlated with the business
cycle.1 The procyclicality of the firm-level investment rate dispersion is noteworthy for at
least two reasons: first, it relates to a growing empirical literature on the time-series dy-
namics of the distribution of micro-level variables; second, as we will argue, it has important
implications for the recent macroeconomic literature on the role of uncertainty or risk shocks
as drivers of the business cycle. The procyclicality of investment dispersion places a robust
and tight upper bound on the aggregate importance of firm-level risk shocks.
Researchers have documented that, across different countries and data sets, the dispersion
of changes in firm- (or plant-) level variables, such as output, productivity, prices and business
forecasts, is robustly countercyclical.2 We find similar comovement patterns for firm-level
output and productivity growth and add firm-level employment growth to the list of variables
the dispersion of which is countercyclical. Since a simple frictionless environment predicts
the dispersions of all decision variables of a firm to comove over the cycle, finding procyclical
investment rate dispersion is interesting and suggestive of an important friction.
We argue that a friction in the capital adjustment technology, nonconvex costs of capital
adjustment, is likely behind this new empirical regularity. First, we show that the strength
of the comovement between investment rate dispersion and the cycle varies with empirical
proxies for the importance of lumpiness in investment. Second, a calibrated heterogeneous-
firm, lumpy investment general equilibrium model with shocks to the idiosyncratic produc-
tivity dispersion, which in the model constitute firm-level risk shocks, can closely match
the observed cyclical comovement of the investment rate dispersion as well as the cyclical
comovement of the dispersion of output and employment growth.
For the calibration of this model, we put front and center another well-established empir-
ical fact about investment at the micro level: the long-run distribution of investment rates
is positively skewed and has excess kurtosis (Caballero et al., 1995). While this previous
research has highlighted the role of nonconvex adjustment costs in shaping the long-run dis-
tribution of investment rates, we add in this paper a fact about the time-series dynamics
of the cross-sectional distribution of micro level investment and again relate it to nonconvex
capital adjustment costs. More generally, we argue that a microfounded business cycle theory
1The literature has documented several related facts: Doms and Dunne (1998) show that the Herfind-ahl index of U.S. plant-level manufacturing investment is positively correlated with aggregate investment.Beaudry et al. (2001) show that cross-sectional investment dispersion in an unbalanced panel of roughly1,000 U.K. manufacturing plants is negatively correlated with conditional inflation volatility. Eisfeldt andRampini (2006) document that capital reallocation in U.S. Compustat data is procyclical.
2See Bachmann and Bayer (2013), Bloom et al. (2012), Doepke et al. (2005), Doepke and Weber (2006),Gourio (2008), Higson et al. (2002, 2004) and Kehrig (2011) for output and/or productivity, Berger andVavra (2011) for prices, and Bachmann et al. (2013b) for business forecasts.
1
can - as we show - successfully speak to the dynamics of more than just the cross-sectional
means of the distributions that underlie macroeconomic aggregates. We view this paper as
a step toward such a research program for the firm sector.3
The procyclicality of the firm-level investment rate dispersion also has implications for
the aggregate business cycle more generally. Arellano et al. (2012), Bloom et al. (2012),
Christiano et al. (2010), Chugh (2012), Gilchrist et al. (2010), Narita (2011), Panousi and
Papanikolaou (2012), Schaal (2011), and Vavra (2012) are examples of recent papers that
have studied the business cycle implications of a time-varying dispersion of firm-specific
variables, often interpreted as and used to calibrate shocks to firm risk, propagated through
various frictions: wait-and-see effects from capital adjustment frictions, financial frictions,
search frictions in the labor market, nominal rigidities and agency problems. We use the
recent literature on risk shocks that are propagated through real options effects as an example
to show that the procyclicality of investment dispersion helps researchers to gauge the power
of risk shocks to generate or alter business cycle fluctuations. To be precise, our model
in the spirit of Bloom (2009) and Bloom et al. (2012) matches the comovement of the
investment rate dispersion with the business cycle only with fluctuations in the cross-sectional
productivity growth dispersion that are in line with direct empirical evidence but lower
than what the previous literature has advocated. More generally, our new fact imposes
a natural overidentifying restriction on any model that uses shocks to firm-level risk that
operate through an investment channel. In a world of noisy micro data, where especially the
direct measurement of firm-level productivity is difficult and invariably assumption-laden,
matching the cyclical behavior of not just the productivity growth dispersion but also the
cyclical behavior of the dispersion of outcome variables, e.g. investment, is a challenge that
models with idiosyncratic risk shocks should meet.
Our primary data source is the Deutsche Bundesbank balance-sheet database of German
firms, USTAN. Hence, the data frequency we observe is annual. This database includes
detailed accounting data that allow us to measure a firm’s value-added, its stock of capital
and its revenue productivity. Another strength of this database is that it covers virtually
the entire nonfinancial private business sector of the German economy, unlike, e.g., the U.S.
Annual Survey of Manufacturing (ASM), and it includes information from many non-traded
medium size companies, unlike COMPUSTAT. This broad coverage permits sample splits
that help us correlate the procyclicality of the investment rate dispersion with empirical
proxies for the importance of lumpiness in investment, namely, industry or firm size. This
makes USTAN uniquely suitable for our purposes. We show nevertheless that the investment
rate dispersion is also procyclical in U.S. (COMPUSTAT) and UK (Cambridge DTI) data.
3Castaneda et al. (1998) have done this for the household side, documenting and explaining the businesscycle dynamics of the U.S. income distribution.
2
Table 1: Cyclicality of Cross-Sectional Moments
Correlation with Cycle
Cross-Sectional Standard Deviation of . . . Fraction of . . .Investment rates 0.45**
Output growth -0.45* Adjusters 0.73***Employment growth -0.50** Spike adjusters 0.61***Invest. rates cond. on spike adj. -0.55***Productivity growth -0.47**
Notes: The left panel refers to the correlation with the cycle of the cross-sectional standard deviations,
linearly detrended, of the investment rate, the log-change of real gross value-added, the net employment
change rate, the investment rate conditional on its absolute value exceeding 20% (spike adjustment), and
the log-change of Solow residuals, all at the firm level. Data are from the Bundesbank’s USTAN database.
We removed firm fixed and 2-digit industry-year effects from each variable. The right panel refers to the
correlation with the cycle of the fraction, linearly detrended, of firms exhibiting an investment rate exceeding
1% (adjusters) and 20% (spike adjusters) in absolute value. The cyclical indicator is the HP(100)-filtered
aggregate real gross value-added in the nonfinancial private business sector, computed from German VGR
(Volkswirtschaftliche Gesamtrechnungen) data. ∗∗∗,∗∗ ,∗ indicate significance at the 1%, 5%, and 10% level,
resulting from an overlapping block bootstrap of four-year windows with 10,000 replications.
Table 1 summarizes our main empirical findings based on USTAN. The firm-level invest-
ment rate dispersion is procyclical. In contrast, the dispersions of output growth, employ-
ment growth, productivity growth and investment rates conditional on an investment spike,
which we define as an investment rate larger than 20% in absolute value, are countercyclical.
Finally, measures of the extensive margin of investment, namely, the fraction of firms with
an investment rate larger than 1% in absolute value and the fraction of spike adjusters are
significantly procyclical.
How do nonconvex capital adjustment costs and a countercyclical idiosyncratic produc-
tivity shock dispersion interact to generate what we find in Table 1? Even abstracting from
risk shocks, nonconvex capital adjustment costs lead to two-step investment rules at the firm
level. Firms first choose whether to adjust or not (extensive margin) and second, conditional
on adjustment, they decide by how much to adjust (intensive margin). The cross-sectional
investment dispersion will in general be a complicated nonlinear function of both steps, but,
as we will show, it is the extensive margin choice that drives the procyclicality in investment
rate dispersions. The difference in the sign of the cyclicalities of investment dispersion and
investment dispersion conditional on an investment spike in Table 1 attests to this fact.
3
To fix ideas, approximate an investment distribution where most investment activity
is concentrated in large lumps by assuming that firms can only decide whether to have an
investment spike or not, i.e., to increase their capital stock by a given and large percentage or
let it depreciate. Under this assumption, both the cross-sectional average and the dispersion
of investment are solely determined by how many firms adjust. Aggregate investment is
increasing in the fraction of firms exhibiting an investment spike, which is why the extensive
margin measures in Table 1 are procyclical. And so is the dispersion of investment if less than
half of the firms exhibit such a spike. Hence, as long as aggregate investment is procyclical
and driven mainly by spike investments, the investment rate dispersion is also procyclical.
Adding a countercyclical productivity shock dispersion lets three additional effects come
into play. With a decrease in the dispersion of productivity shocks, on the one hand firms will
tend to adjust less often, as they simply move more slowly over their adjustment triggers. If
low dispersion is concentrated in booms, then this volatility effect will tend to counteract the
procyclicality of the extensive margin and the dispersion of investment. On the other hand,
to the extent that a decrease in dispersion also constitutes a decrease in firm-level risk, firms
see a decline in the option value of waiting, narrow their adjustment triggers, and thus adjust
their capital stock more frequently. With a countercyclical productivity shock dispersion,
this real options effect will tend to strengthen the procyclicality of the extensive margin and
the dispersion of investment. Finally, there is the intensive margin effect that goes in the
same direction as the volatility effect. Conditional on adjustment, costs are sunk and firms
behave similarly to a frictionless setup. Hence conditional on adjustment, the distribution
of investment rates follows the distribution of shocks, i.e., the conditional investment rate
dispersion is negatively correlated with the cycle. The same argument holds for all firm-level
decision variables that are not subject to large fixed costs of adjustment.
After expounding on the empirical findings in Section I, we ask in Sections II-IV whether
the qualitative intuition described here holds up in a fully specified heterogeneous-firm real
business cycle model that features both the extensive and the intensive margins of capital
adjustment as well as shocks to aggregate productivity and countercyclical dispersion of firm-
level productivity shocks. We find that indeed such a model, whose adjustment costs are
calibrated to match the skewness and kurtosis of the long-run investment rate distribution,
can quantitatively match the empirical procyclicality of the investment rate dispersion. We
also show that small risk shocks to firm-level productivity are necessary for this result, in
that the volatility and the intensive margin effect must be sufficiently yet not too strong to
be quantitatively consistent with the procyclicality of the dispersion of investment rates.
4
I The Facts
We showed in the Introduction that the dispersion of firm-level investment rates is procyclical
despite the dispersion of productivity, output and employment growth being countercyclical.
In this section, we first fill in some information about our primary data source (USTAN),
which is complemented by a more detailed description in Appendix A. We then link the
procyclicality of investment dispersion to proxies for the lumpiness of investment, i.e., we
identify firms where investment lumpiness should be more prevalent and show that the
investment rate dispersion is more procyclical for those firms. We then show that investment
rate dispersions are also procyclical in the U.S. and the UK and conclude with a summary
of the robustness checks that we present in detail in Appendix B.
I.A Data and Sample Selection
Our primary data source is the Deutsche Bundesbank balance-sheet database of German
firms, USTAN. USTAN is an annual private-sector, firm-level data set that allows us to
make use of 26 years of data (1973-1998), with cross-sections that have, on average, over
30,000 firms per year. For the U.S. and UK evidence we use the COMPUSTAT sample
from 1970-1994 and the Cambridge DTI database for 1978-1990, respectively. After the
data undergo a similar data treatment to that for USTAN, the former covers roughly 2,150
firms per year and the latter only 850.4 As usual, we compute economic capital stocks by
a perpetual inventory method from balance-sheet data (see Appendix A.4 for details). We
exploit the fact that all three data sets provide information for capital disaggregated into
structures and equipment. This makes our measure of the economic capital stock robust to
heterogeneity in capital portfolios, when, for instance, some firms have a larger fraction of
their capital invested in structures than others.5 The size of USTAN and its broad coverage
in terms of ownership, firm size and industry allow us to study the cyclicality of investment
rate dispersions in various sample splits that are meant to capture putative differences in
the relevance of investment lumpiness.
USTAN is a byproduct of the Bundesbank’s rediscounting activities. The Bundesbank
had to assess the creditworthiness of all parties backing promissory notes or bills of exchange
put up for rediscounting (i.e., as collateral for overnight lending). It implemented this
regulation by requiring balance-sheet data of all parties involved, and the data were then
4When cross-sectional dispersions are concerned, Davis et al. (2006) show that studying only publiclytraded firms (COMPUSTAT) can lead to misleading conclusions, which is why we view USTAN as verysuitable for our purposes.
5The heterogeneity of capital portfolios is also the reason why we restrict the COMPUSTAT sample to1970-1994. After 1994 no separate information for structures and equipment is available there.
5
collected and archived (see Appendix A.1, Stoess (2001) and von Kalckreuth (2003) for
details).
From the original USTAN data, we select firms that report information on payroll, gross
value-added and capital stocks, and for which we have at least five observations in first
differences. Moreover, we drop outliers and observations from East German firms to avoid
a break in the series in 1990 (see Appendix A.2 for details). This leaves us with a sample of
854,105 firm-year observations from 72,853 different firms, i.e., on average a firm is observed
in the sample for 11.7 years. The average number of firms in the cross-section of any given
year is 32,850. The resulting sample covers roughly 70% of the West German real gross
value-added in the nonfinancial private business sector and 50% of its employment.6
Throughout the paper, we follow Bloom (2009) and define the investment rates of a firm
j at time t as ij,t =Ij,t
0.5(kj,t+kj,t+1).7 Investment rates exhibit cross-sectional skewness (2.19)
and kurtosis (20.04), which, according to Caballero et al. (1995), is a strong indication of
investment lumpiness at the firm level. Table 20 in Appendix A.6 shows that the investment
histograms for USTAN and the manufacturing sector in USTAN look very similar to the
investment histogram for the U.S. in Cooper and Haltiwanger (2006).
For firm-level employment growth rates we use the symmetric adjustment rate definition
proposed in Davis et al. (1996),∆nj,t
0.5∗(nj,t−1+nj,t). Firm-level productivity and output growth
rates are simple log-differences of, respectively, Solow residuals and real gross value-added,
for which we deflate the balance-sheet item nominal gross value-added by the price index for
gross value-added from German national accounting data (VGR).8 To focus on idiosyncratic
changes that do not capture differences in industry-specific responses to aggregate shocks or
ex-ante firm heterogeneity, firm fixed and industry-year effects are removed from investment
rates, as well as from the employment, output and productivity growth rates.
6Throughout we will refer to Agriculture, Mining and Energy, Manufacturing, Construction, Trade andTransportation and Communication collectively as the nonfinancial private business sector (NFPBS).
7Spike adjusters are defined relative to this investment rate definition, i.e., ij,t < −20% or ij,t > 20%.Strictly speaking, the literature, e.g., Cooper and Haltiwanger (2006) and Gourio and Kashyap (2007),
has used the 20% threshold with respect toIj,tkj,t
, but we show in Appendix A.6 that the investment rate
histograms in USTAN look similar for either definition of the investment rate.8To compute firm-level Solow residuals, we start, in accordance with the model in Section II, from
a firm-level Cobb-Douglas production function: yj,t = exp(zt + εj,t)kθj,tn
νj,t, where ε is firm-specific and z
aggregate log productivity. We assume that labor input n is immediately productive, whereas capital k is pre-determined and inherited from the last period. This difference is reflected in the different timing conventionin the definitions of the investment and employment adjustment rates. We estimate the output elasticitiesof the production factors, ν and θ, as median factor expenditure shares over gross value-added within eachindustry. Measured Solow residuals will likely reflect true firm productivity with some error. We take thisinto account and perform a measurement error correction, estimating the size of the measurement error bycomparing the variances of one- and two-year Solow residual growth rates. See for details Appendices A.7and A.8.
6
I.B Procyclicality of Investment Rate Dispersions and Proxies for
Lumpy Investment
The literature has typically focused on the manufacturing sector to find evidence for noncon-
vex adjustment technologies (see Doms and Dunne (1998), Caballero et al. (1995), Caballero
and Engel (1999), Cooper and Haltiwanger (2006), and Gourio and Kashyap (2007)), and for
good reason: manufacturing is where heavy-duty machinery needs to be installed and large
production halls need to be built, which may lead to disruptions in the production process.
Indeed, Table 2 shows that in manufacturing and also in construction, the correlation of
the investment rate dispersion with the industry cycle is particularly strong and statistically
significant.
Table 2: Cyclicality of Cross-Sectional Investment Rate Dispersion - By One-Digit Industries
Correlation of Real Gross Value-Added with std(ij,t)
All Industries 0.45∗∗
Primary Sector Secondary Sector Tertiary Sector
Agriculture −0.19 Manufacturing 0.48∗∗∗ Trade 0.21Mining & 0.04 Construction 0.44∗∗ Transport & 0.40∗∗
Energy Communication
Notes: See notes to Table 1. The table displays correlation coefficients with the cyclical component of
aggregate real gross value-added of the nonfinancial private business sector in the first row, thereafter with
the real gross value-added of the corresponding one-digit industry.
Another dimension that is likely correlated with the relevance of adjustment frictions is
firm size. Larger firms may partially outgrow fixed adjustment costs or can smooth the effects
of nonconvex capital adjustment costs and the extensive margin over several production units.
In Table 3 we see that the procyclicality of the investment rate dispersion is falling in firm
size. The very large firms, in contrast to the small ones, have an almost acyclical investment
dispersion. This distinction is statistically significant in the sense that if size is measured in
terms of employment or value-added, neither the point estimate for the smallest size class
lies in the [5%, 95%]−band of the largest size class nor vice versa.
Another way to see how the extensive margin of investment and the dispersion of firm-
level productivity growth interact to generate a procyclical investment rate dispersion is to
7
Table 3: Cyclicality of Cross-Sectional Investment Dispersion - By Firm Size
Size Class / Criterion Employment Value-Added Capital
Smallest 25% 0.58∗∗∗ 0.60∗∗∗ 0.39∗∗
25% to 50% 0.46∗∗ 0.47∗∗ 0.42∗∗
50% to 75% 0.37∗ 0.33 0.39∗
Largest 25% 0.19 0.22 0.40∗∗
Largest 5% 0.05 0.05 0.18
Notes: See notes to Tables 1 and 2. Just as for the aggregate numbers in Table 1, we use the cyclical
component of the aggregate output of the private nonfinancial business sector as the cyclical indicator.
exploit the sectoral information in our data. We disaggregate the data by years and 14 two-
digit industries and then regress in a pooled OLS regression the dispersion of investment
rates on the dispersion of firm-level productivity growth and the fraction of (spike) adjusters
in a given industry and year. Table 4 shows that the larger the dispersion of shocks is and
the more frequent investment activities in an industry-year are, the larger is the disper-
sion of investment rates. Importantly, conditional on the fraction of (spike) adjusters, i.e.,
the extensive margin of investment, investment dispersion and the dispersion of firm-level
productivity growth comove positively. Since the productivity growth dispersion is coun-
tercyclical, the procyclicality of the dispersion of investment rates has to be driven by the
procyclicality of the cross-sectional investment frequency. Put differently, the fluctuations
in firm-level risk cannot be too large so as to undo the extensive margin effect.
Table 4: Evidence from Disaggregation by Two-Digit Industry and Year
Regression of std(ij,t) on ...(a) (b)
Fraction of Adjusters .23∗∗∗ Fraction of Spike Adjusters .28∗∗∗
std(∆εj,t) .37∗∗∗ std(∆εj,t) .20∗∗∗
Notes: The table displays the estimated coefficients of a pooled OLS regression of the cross-sectional invest-
ment rate dispersion for each two-digit industry and year on the fraction of (spike) adjusters in that industry
and year and the dispersion of idiosyncratic productivity shocks. All data have been linearly detrended at
the industry level. See Table 15 in Appendix A.3 for more details on the two-digit industries in USTAN.
8
Table 5: Evidence from U.S. and UK data
Correlation of std(ij,t) with HP(100)-Y
U.S. (COMPUSTAT) 0.60∗∗∗ UK (Cambridge DTI) 0.45∗∗
Notes: See notes to Table 1. The cyclical indicator HP(100)-Y refers to the cyclical component of aggregate
real gross value-added of the nonfinancial private business sector from NIPA data for the U.S. and, because
the DTI has a high fraction of manufacturing firms, to the manufacturing production index for the UK.
Collectively, the evidence in this section at least suggests that nonconvex capital adjust-
ment costs play a role in explaining procyclical investment dispersion.
To conclude, we show that the procyclicality of the investment rate dispersion is robust
across different data sets. We use the U.S. COMPUSTAT data and the Cambridge DTI data
to document procyclical investment dispersions for the U.S. and the UK. Wherever possible,
we treat the data analogously to the way we treated the USTAN data. As cross-sections are
much smaller in the U.S. and UK data, the sample splits that we do for USTAN are not
possible. Again we select firms from the nonfinancial private business sector and correlate
their investment rate dispersions with the corresponding aggregate gross real value-added.
Table 5 displays the results.
I.C Robustness
How robust is the procyclicality of the investment rate dispersion? Potential issues for
robustness are: our measure of the cycle, our measure of dispersion, and the representative-
ness of the sample both cross-sectionally and in the time-series dimension. We establish the
robustness of our findings in all these dimensions in Appendix B in Tables 21, 22, and 23.
We check robustness to alternative measures of the cycle (HP-filter parameters, aggre-
gate variables indicating the cycle, detrending of the dispersion series, etc.), to looking at
dynamic correlations, to excluding the two most extreme investment dispersion years, and
to excluding the post-reunification period. Moreover, we check robustness with respect to
sample composition. We study a sample where we include firms only three years after they
entered the sample the first time, to make sure our result is not driven by firm entry. We also
look at firms that are stable in the sample so that cyclicality is not driven by the systematic
exit of firms. Finally, we try different criteria for excluding outliers, using a non-centralized
definition of investment rates, alternative ways of treating movements in the price of capital
goods, and replacing the standard deviation as our measure of dispersion with the interquar-
tile range.
9
II The Model
We follow closely Khan and Thomas (2008) and Bachmann et al. (2013a). The main depar-
ture from both papers is the introduction of a second aggregate shock, namely, time-varying
idiosyncratic productivity risk.
II.A Firms
The economy consists of a unit mass of small firms. There is one commodity in the economy
that can be consumed or invested. Each firm produces this commodity, employing labor (n)
and its pre-determined capital stock (k), according to the following Cobb-Douglas decreasing-
returns-to-scale production function:
y = exp(z + ε)kθnν ; with θ, ν > 0 and θ + ν < 1, (1)
where z and ε are log aggregate and log idiosyncratic revenue productivity, respectively.
The idiosyncratic log productivity process is first-order Markov with autocorrelation ρε
and time-varying conditional standard deviation, σ(ε). We assume two exogenous aggregate
states (z, s), which evolve jointly according to an unrestricted VAR(1) process, with normal
innovations u that have zero mean and covariance Ω:9(z′
s′
)= %A
(z
s
)+ u, cov(u) = Ω. (2)
In line with the production function (1) z is the trend deviation of the natural logarithm
of aggregate productivity, while s drives the dispersion of idiosyncratic productivity shocks,
which is given by σ(ε) = σσs+ σ(ε), where σ(ε) denotes the steady-state standard deviation
of the innovations to idiosyncratic productivity, and σσ scales the size of the fluctuations in
σ(ε). The shocks to the exogenous aggregate states, u, and idiosyncratic productivity shocks
are independent. Idiosyncratic productivity shocks are independent across productive units.
We do not impose any restrictions on Ω or %A ∈ R2×2.
The trend growth rate of aggregate productivity is (1−θ)(γ−1), so that aggregate output
and capital grow at rate γ − 1 along the balanced growth path. From now on we will work
with k and y (and later aggregate consumption, C) in efficiency units.
We model employment as freely adjustable but assume that capital adjustment is costly.
Each period, a firm draws its current cost of capital adjustment, 0 ≤ ξ ≤ ξ, which is denom-
inated in units of labor, from a time-invariant distribution, G. G is a uniform distribution
on [0, ξ], common to all firms. Draws are independent across firms and over time.
9Curdia and Reis (2011) recently pointed to using correlated shocks for understanding business cycles.
10
Upon investment the firm incurs fixed costs ωξ, where ω is the current real wage. Capital
depreciates at rate δ. We denote the firms’ distribution over (ε, k) by µ. Thus,(z, s, µ
)constitutes the current aggregate state and µ evolves according to the law of motion µ′ =
Γ(z, s, µ
), which firms take as given.
Next we describe the dynamic programming problem of a firm. Following Khan and
Thomas (2008), we state this problem in terms of utils of the representative household (rather
than physical units) and denote the marginal utility of consumption by p = p(z, s, µ
). This is
the kernel that firms use to price output streams. Also, given the i.i.d. nature of adjustment
costs, continuation values can be expressed without future adjustment costs.
Let V 1(ε, k, ξ; z, s, µ
)denote the expected discounted value - in utils - of a firm that is
in idiosyncratic state (ε, k, ξ), given the aggregate state(z, s, µ
). Then the firm’s expected
value prior to the realization of the adjustment cost is:
V 0(ε, k; z, s, µ
)=
∫ ξ
0
V 1(ε, k, ξ; z, s, µ
)G(dξ). (3)
With this notation the dynamic programming problem becomes:
V 1(ε, k, ξ; z, s, µ
)= max
nCF + max(Vno adj,max
k′[−AC + Vadj]), (4)
where CF denotes the firm’s flow value, Vno adj the firm’s continuation value if it chooses
inaction and does not adjust, and Vadj the continuation value, net of adjustment costs AC,
if the firm adjusts its capital stock. That is:
CF = [exp(z + ε)kθnν − ω(z, s, µ)n]p(z, s, µ), (5a)
Vno adj = βE[V 0(ε′, (1− δ)k/γ; z′, s′, µ′)], (5b)
AC = ξω(z, s, µ)p(z, s, µ), (5c)
Vadj = −(γk′ − (1− δ)k
)p(z, s, µ) + βE[V 0(ε′, k′; z′, s′), µ′)], (5d)
where both expectation operators average over the next period’s realizations of the aggregate
and idiosyncratic shocks, conditional on this period’s values. The discount factor, β, reflects
the time preferences of the representative household.
Taking as given ω(z, s, µ) and p(z, s, µ), and the law of motion µ′ = Γ(z, s′, µ), the firm
chooses optimal labor demand, whether to adjust its capital stock at the end of the period,
and the optimal capital stock, conditional on adjustment. This leads to policy functions:
N = N(ε, k; z, s, µ) and K = K(ε, k, ξ; z, s, µ). Since capital is pre-determined, the optimal
employment decision is independent of the current adjustment cost draw.
11
II.B Households
We assume a continuum of identical households. They have a standard felicity function in
consumption and labor:
U(C,Nh) = logC − ANh, (6)
where C denotes consumption and Nh the households’ labor supply. Households maximize
the expected present discounted value of the above felicity function, yielding:
p(z, s, µ
)≡ UC(C,Nh) =
1
C(z, s, µ
) , and ω(z, s, µ
)= −UN (C,Nh)
p(z,s,µ) = A
p(z,s,µ) . (7)
II.C Recursive Equilibrium
A recursive competitive equilibrium for this economy is a set of functions(ω, p, V 1, N,K,C,Nh,Γ
),
that satisfy
1. Firm optimality : Taking ω, p and Γ as given, V 1(ε, k, ξ; z, s, µ) solves (4) and the
corresponding policy functions are N(ε, k; z, s, µ) and K(ε, k, ξ; z, s, µ).
2. Household optimality : Taking ω and p as given, the household’s consumption and labor
supply satisfy (7).
3. Commodity market clearing :
C(z, s, µ) =
∫exp(z+ε)kθN(ε, k; z, s, µ)νdµ −
∫ ∫ ξ
0
[γK(ε, k, ξ; z, s, µ)−(1−δ)k]dGdµ.
4. Labor market clearing :
Nh(z, s, µ) =
∫N(ε, k; z, s, µ)dµ +
∫ ∫ ξ
0
ξJ(γK(ε, k, ξ; z, s, µ)− (1− δ)k
)dGdµ,
where J (x) = 0, if x = 0 and 1, otherwise.
5. Model consistent dynamics : The evolution of the cross-section that characterizes the
economy, µ′ = Γ(z, s, µ), is induced by K(ε, k, ξ; z, s, µ) and the exogenous processes
for z, s as well as ε.
Conditions 1, 2, 3 and 4 define an equilibrium given Γ, while step 5 specifies the equilib-
rium condition for Γ.
12
II.D Solution
It is well-known that (4) is not computable, because µ is infinite dimensional. We follow
Krusell and Smith (1997, 1998) and approximate the distribution, µ, by a finite set of its
moments, and its evolution, Γ, by a simple log-linear rule. As usual, we include aggregate
capital holdings, k. We find that adding the unconditional cross-sectional standard deviation
of the natural logarithm of the level of idiosyncratic productivity, std(ε), not only improves
the fit of the Krusell-Smith rules but it also matters for our economic results (see Appendix C
for details).10 This is owing to the now time-varying nature of the distribution of idiosyncratic
productivity. We surmise that this is a general insight and that simple Krusell-Smith rules are
likely inappropriate in models with firm-level risk shocks. In the same vein, we approximate
the equilibrium pricing function by a log-linear rule:
log k′ =ak(z, s)
+ bk(z, s)
log k + ck(z, s)
log std(ε), (8a)
log p =ap(z, s)
+ bp(z, s)
log k + cp(z, s)
log std(ε). (8b)
Given (7), we do not have to specify a rule for the real wage. We posit the rules (8a)–(8b)
and check that in equilibrium they yield a good fit to the actual law of motion.11
Substituting k and std(ε) for µ and using (8a)–(8b), (4) becomes a computable dynamic
programming problem with corresponding policy functions N = N(ε, k; z, s, k,
std(ε))
and K = K(ε, k, ξ; z, s, k, std(ε)
). We solve this problem by value function itera-
tion on V 0 and apply multivariate spline techniques that allow for a continuous choice of
capital when the firm adjusts.
With these policy functions, we can simulate a model economy without imposing the
equilibrium pricing rule (8b). Rather, we impose market-clearing conditions and solve for
the pricing kernel at every point in time of the simulation. This generates a time series
of pt and kt endogenously, on which the assumed rules (8a)–(8b) can be updated with
a simple OLS regression. The procedure stops when the updated coefficients ak(z, s′
)to
cp(z, s′
)are sufficiently close to the previous ones.
III Calibration
The model frequency is annual, which corresponds to the data frequency in USTAN. Some
model parameters are directly calculated or estimated from VGR and/or USTAN data (such
10For a similar insight, see Zhang (2005).11std(ε) is a function of std(ε−1) and σ(ε). Further details on the numerical solution method and on the
quality of the approximation are available in Appendix C.
13
Table 6: Model Parameters: Baseline Calibration
Parameter Value Calibrated from / to
discount factor β 0.97 real interest rate on corporate bonds: 4.6%
disutility of labor A 2 average time spent at work: 1/3
depreciation rate δ 0.094 VGR Data: depreciation rate
long-run growth factor γ 1.014 VGR Data: aggregate investment rate
output elasticity of labor ν 0.5565 USTAN: value-added share
output elasticity of capital θ 0.2075 USTAN: value-added share
time-average idiosyncratic risk σ(ε) 0.0905 USTAN: Solow residual growth dispersion
autocorrelation of idiosyncratic produc’ty ρε 0.9675 USTAN: Solow residual growth
Joint process ofaggregate productivity and %A,Ω see below VGR Data: Solow residuals
volatility of idiosyncratic risk USTAN: Solow residual growth dispersion
Scaling of risk fluctuations σσ 1 Normalization
adjustment cost parameter ξ 0.2 USTAN: investment rate
skewness and kurtosis
as the depreciation rate, the output elasticities and the parameters of the aggregate and
idiosyncratic driving processes). The remaining parameters are jointly calibrated to match
the real interest rate, the average time spent at work and the aggregate investment rate
in the German nonfinancial private business sector as well as the skewness and kurtosis of
the firm-level investment rate in the USTAN data. Table 6 gives an overview of the model
parameters, their values and the data sources. Importantly, the cyclicality of the investment
rate dispersion is not targeted, nor are the cyclicalities of the fraction of spike adjusters, the
output growth dispersion and the employment growth dispersion.
III.A Technology and Preference Parameters
We take the depreciation rate, δ = 0.094, directly from German national accounting (VGR)
data for the nonfinancial private business sector. Given this depreciation rate, γ = 1.014
matches the time-average aggregate investment rate in the nonfinancial private business
sector: 0.108.12 Since the average real interest rate in Germany over the period 1973-1998
12γ = 1.014 is also consistent with German long-run growth rates.
14
was 4.6%, we obtain for the discount factor β = 0.97.13 The disutility of work parameter,
A, is chosen to generate an average time spent at work of 0.33: A = 2.
We set the output elasticities of labor and capital to ν = 0.5565 and θ = 0.2075, respec-
tively, which correspond to the measured median labor and capital shares in manufacturing
in the USTAN database. Our model simulations show that value-added shares are good
estimators of the output elasticities even in the presence of fixed adjustment costs and out-
perform other estimators of the production function; see Appendix A.7 for details.
III.B Idiosyncratic Shocks
We calibrate the standard deviation of idiosyncratic productivity shocks to σ(ε) = 0.0905,
which we obtain from measured firm-level Solow residual growth in USTAN cleansed of
measurement error. We set ρε = 0.9675, which we estimate again from measured productivity
in USTAN (details are available in Appendix A.8). This process is discretized on a 19−state-
grid, using Tauchen’s (1986) procedure with a mixture of two Gaussian normals to capture
above-Gaussian kurtosis - 4.4480 on average - in idiosyncratic productivity shocks (details
are available in Appendix C). Heteroskedasticity in the idiosyncratic productivity process
is modeled with time-varying transition matrices between idiosyncratic productivity states,
where the matrices correspond to different values of σ(ε).
III.C Aggregate Shocks
To calibrate the parameters of the two-state aggregate shock process, we estimate a bivariate,
unrestricted VAR with the linearly detrended natural logarithm of the aggregate Solow
residual14 and the linearly detrended σ(ε)-process, i.e., the process for the standard deviation
of the innovations to idiosyncratic productivity, from the USTAN data, where we normalize
σσ = 1 for the baseline model. The estimated parameters of this VAR are:15
%A =
(0.2791 −1.3439∗∗
0.1059∗∗ 0.8072∗∗∗
)Ω =
(0.0115∗∗∗ −0.5459∗∗∗
−0.5459∗∗∗ 0.0036∗∗∗
)(9)
Importantly, both the negative contemporaneous correlation and the negative coefficient of
firm risk on future TFP are significant. This process is discretized on a [5 × 5]−state grid,
using a bivariate analog of Tauchen’s procedure.
13We calculate the real interest rate as the return on corporate bonds minus the ex-post inflation rate.14We use ν = 0.5565 and θ = 0.2075 in these calculations.15With a slight abuse of notation, but for the sake of readability, Ω has standard deviations on the main
diagonal and correlations on the off diagonal. ** and *** denote the usual significance levels. Notice the highpersistence in the σ(ε)-process. Today’s idiosyncratic productivity shock dispersion has strong predictivepower about tomorrow’s idiosyncratic productivity shock dispersion and thus reflects risk shock.
15
Table 7: Calibration of Adjustment Costs - ξ
ξ Skewness Kurtosis Ψ(ξ)0.00 -0.01 3.61 18.930.01 1.13 5.83 8.980.10 2.45 10.88 2.900.20 (BL) 2.90 13.05 2.660.30 3.17 14.46 3.010.50 3.51 16.40 4.011.00 3.96 19.37 6.51
Notes: ‘BL’ denotes the baseline calibration. Skewness and kurtosis refer to the time-average of the corre-
sponding cross-sectional moments of firm-level investment rates. The fourth column displays the value of Ψ,
the precision-weighted Euclidean distance of the model’s cross-sectional skewness and kurtosis of investment
rates to their data counterparts.
III.D Adjustment Costs
The distribution of firm-level investment rates exhibits both substantial positive skewness
– 2.1920 – as well as kurtosis – 20.0355. Caballero et al. (1995) document a similar fact
for U.S. manufacturing plants. They also argue that nonconvex capital adjustment costs
are an important ingredient for explaining such a strongly non-Gaussian distribution, given
a close-to-Gaussian firm-level shock process. With fixed adjustment costs, firms have an
incentive to lump their investment activity together over time in order to economize on
these adjustment costs. Therefore, typical capital adjustments are large, which creates excess
kurtosis. Making use of depreciation, firms can adjust their capital stock downward without
paying adjustment costs. This makes negative investments less likely and hence leads to
positive skewness in firm-level investment rates. We therefore use the skewness and kurtosis
of firm-level investment rates to identify ξ.
Since, as a practical matter, the adjustment cost parameter, ξ, hardly impacts long-
run variables, such as the average real interest rate, the average time spent at work or the
average aggregate investment rate in the model, it is convenient to proceed as follows: given
the following set of parameters β, δ, γ, A, ν, θ, σ(ε), ρε, %A,Ω, σσ, we find ξ by minimizing the
Euclidean distance, Ψ(ξ), between the time-average firm-level investment rate skewness and
kurtosis produced by the model and the data. To take into account the different precision at
which we estimate skewness and kurtosis, we weigh both with the inverse of their time-series
standard deviation. Table 7 shows that ξ is indeed identified in this calibration strategy,
as cross-sectional skewness and kurtosis of the firm-level investment rates are monotonically
increasing in ξ. The minimum of Ψ is achieved at ξ = 0.2, our baseline.
16
Before investigating what this calibration implies for the correlation of the dispersion of
various firm-level activity variables with aggregate economic activity, it is also instructive
to see what it entails for other statistics related to nonconvex capital adjustment costs,
which have not been targeted here. The average adjustment costs conditional on adjustment
amount to roughly 11% as a fraction of annual firm-level output, which is at the lower end
of estimates from the U.S. (see Bloom (2009), Table IV, for an overview). Moreover, the
baseline model implies a fraction of spike adjusters of 11.3%, i.e., firms with an investment
rate that is larger than 20% in absolute value, which is well in line with the 13.4% in
the USTAN data. Finally, our model produces basically zero autocorrelation of firm-level
investment rates (-0.05), compared to −0.03 in the USTAN data, a typical feature of lumpy
investment at the micro-level.16
IV Results
IV.A Baseline Results
Can a general equilibrium model with standard aggregate productivity shocks, persistent
idiosyncratic productivity shocks, countercyclical aggregate shocks to their dispersion and
fixed capital adjustment costs, calibrated to the long-run non-Gaussianity of the investment
rate distribution, reproduce the non-targeted cyclicality of the cross-sectional dispersion of
firm-level investment rates, output growth and employment growth and the cyclicality of the
extensive and intensive margins of spike investment? Table 8 says yes.
That the model can closely match the cyclicality of the investment rate dispersion is an
example of the larger premise of this paper: that cross-sectional dynamics are an important
aspect of the data that heterogeneous firm models should address. With quantitatively
realistic shocks to the dispersion of firm-level Solow residuals, it is one level of adjustment
costs that makes the model jointly consistent with the (targeted) time-average skewness and
kurtosis of the investment rate distribution – two statistics closely related to the relevance of
nonconvexities at the micro-level – and the time-series correlation between the cross-sectional
standard deviation of investment rates and output (not targeted).
Let us reiterate the intuition for the results in Table 8. In a world with aggregate
productivity shocks only, and firms merely having to make a decision about whether to
realize an investment spike or not, both the fraction of spike adjusters and the dispersion
of investment rates are procyclical, as long as spike adjustment is sufficiently infrequent:
16Cooper and Haltiwanger (2006) found an autocorrelation of plant-level investment rates of 0.06 in theU.S. LRD data and use this number as one of the characteristic moments in their GMM procedure to identify(non-)convex capital adjustment costs.
17
Table 8: Cyclicality of Cross-Sectional Dispersions and the Margins of In-vestment - Baseline Model
Correlation with the cycle
Cross-sectional Moment Model Data
std(ij,t) 0.53 0.45Fraction of spike adjusters 0.63 0.61
std(∆ log yj,t) -0.36 -0.45
std(∆nj,t
0.5∗(nj,t−1+nj,t)) -0.38 -0.50
std(ij,t) conditional on spike adjustment -0.74 -0.55
Notes: See notes to Table 1. The table displays correlation coefficients with HP(100)-filtered aggregate
output. The column ‘Model’ refers to the correlation coefficients from a simulation of the baseline model.
aggregate gross investment is given by λκ, where λ is the fraction of spike adjusters and
κ the size of the investment spike. Investment dispersion in such an economy would be
λ(1− λ)κ2, which is increasing in the fraction of spike adjusters as long as λ ≤ 0.5, which is
the case in the data. This intuition carries over into a more realistic economy where there is
an intensive margin of investment that is nevertheless of secondary (to the extensive margin)
importance for aggregate investment, again as in the data.
With a countercyclical productivity shock dispersion, the real options effect of higher
firm-level risk will strengthen the procyclicality of the investment rate dispersion, as lower
risk increases the fraction of spike adjusters. In contrast, both the volatility effect – lower risk
decreases the probability that firms hit their adjustment triggers – and the intensive margin
effect counteract the procyclicality of the investment rate dispersion. The volatility and the
intensive margin effects appear to dominate the real-options effect, which is suggested by
the fact that for both the investment rate dispersion and for the fraction of spike adjusters
we see positive correlations with the cycle that are substantially smaller than those same
correlations in a model without risk shocks (see Table 9 below).
Going back to Table 8, the countercyclicality of the output and employment growth
dispersion follows directly from the countercyclical productivity shock dispersion, as both
output and employment growth are simple functions of productivity growth and the growth
of the capital stock of a firm, essentially the investment rate one period ago. Ignoring the
cross-sectional covariances and their time-series behavior, the impact of which is small with
idiosyncratic productivity being close to a random walk, the cyclicality of the output and
employment growth dispersion is then simply a function of the cyclicality of the dispersions
18
Table 9: Adjustment Costs and the Cyclicality of Cross-Sectional Variables
with 2nd moment shocks w/o 2nd moment shocks
ξ std(ij,t) std(∆ log yj,t) Fraction of std(ij,t) std(∆ log yj,t) Fraction ofspike adjusters spike adjusters
0 -0.41 -0.46 -0.41 - - -0.010.01 -0.23 -0.44 0.17 0.85 0.11 0.690.1 0.23 -0.39 0.34 0.88 0.20 0.880.2 0.53 -0.36 0.63 0.88 0.21 0.910.3 0.68 -0.35 0.76 0.89 0.23 0.910.5 0.82 -0.32 0.86 0.89 0.18 0.921 0.91 -0.29 0.92 0.90 0.17 0.92
Notes: See notes to Table 8. ‘with 2nd moment shocks’ refers to a simulation with aggregate productivity
shocks and shocks to the dispersion of firm-level Solow residuals, as specified in equation (9). ‘w/o 2nd
moment shocks’ refers to a simulation with only aggregate productivity shocks, where %A = 0.5223 and
Ω = 0.0121. Note that in this case, with ξ = 0, std(ij,t) and std(∆ log yj,t) are constant, which means that
their correlation coefficients with output are not defined.
of productivity shocks and investment rates. Since as usual with Cobb-Douglas production
functions the coefficient on factor growth is an order of magnitude smaller than that on
productivity growth, the cyclicality of the productivity shock dispersion will dominate.
To further investigate the mechanism behind the procyclical investment rate dispersion,
Table 9 displays the cyclicality of the investment rate and output growth dispersions as well
as the cyclicality of the fraction of spike adjusters that the model generates for various levels
of adjustment costs, both with and without second moment shocks.17
Two findings are important:
1. The right panel of Table 9 shows that, without second moment shocks, neither the
procyclicality of the investment dispersion, the procyclicality of the fraction of lumpy
adjusters, nor the countercyclicality of the output growth dispersion can be quanti-
tatively replicated. Already a very small nonconvex capital adjustment cost factor
generates procyclical investment dispersion. The model overshoots the number in the
data considerably. Also, without countercyclical second moment shocks, the dispersion
of value-added growth is slightly procylical. This follows immediately from the consid-
erations above: without time-varying dispersion of productivity growth, the cyclicality
17The cyclicality of the employment growth dispersion, for space reasons not shown in Table 9, behavessimilarly to that of the output growth dispersion.
19
of the dispersion of output growth is determined by the cyclicality of (lagged) invest-
ment rates. Hence, countercyclical second moment shocks play an important role in
understanding cross-sectional firm dynamics.
2. In the presence of countercyclical second moment shocks (left panel), in the frictionless
case, ξ = 0, the dispersions of investment and output growth merely mirror the coun-
tercyclicality of the dispersion of the idiosyncratic driving force. Because it is more
likely to observe an investment spike when the dispersion of shocks goes up (at ξ = 0
the volatility and the intensive margin effect are the same), the fraction of spike ad-
justers is also countercyclical. Increasing the fixed adjustment costs leads to a gradual
increase in the procyclicality of the investment rate dispersion and the procyclicality
of the fraction of spike adjusters. The real options effect of countercyclical risk shocks
becomes stronger; more firms invest in low-risk times, i.e., in booms. At ξ = 1 this
effect eventually starts to dominate the volatility effect, and the model with second
moment shocks has a more procyclical investment dispersion than the one without.
Table 10 shows how our baseline results change with more volatile firm-level risk. To
this end, we double and quadruple, respectively, the scaling parameter, σσ, in the definition
of σ(ε), and re-estimate (2). This means that we double/quadruple the time-series coeffi-
cient of variation of firm-level risk. Given that the data suggest a correlated shock process
between aggregate productivity and firm-level risk (see equation 9), it is important to keep
the information structure on aggregate productivity the same and isolate the pure effect of
time-varying firm-level risk. This is conveniently done by setting σσ = 1, 2, or 4.
Table 10: Volatility of Risk and Cyclicality of Investment Disper-sion/Extensive Margin
Volatility of std(∆εj,t)
Baseline Double QuadrupleCross-sectional Moment σσ = 1 σσ = 2 σσ = 4
std(ij,t) 0.53 -0.09 -0.37std(∆ log yj,t) -0.36 -0.42 -0.45Fraction of spike adjusters 0.63 -0.10 -0.41
Notes: See notes to Table 8.
20
Increasing the time-series volatility of countercyclical firm-level risk leads, counterfactu-
ally, to acyclical or even countercyclical behavior of the investment rate dispersion. The
table also shows that the dispersion of productivity growth affects the investment dispersion
not only through the intensive margin of investment – conditional on adjustment, invest-
ment becomes more disperse when shocks are more disperse – but also through the extensive
margin: with more volatile firm-level risk shocks, the volatility effect makes the fraction of
spike adjusters countercyclical – more – not less – firms adjust when σ(ε) is high.
In summary, Table 10 shows that only our baseline specification with relatively small
fluctuations in idiosyncratic productivity growth dispersions matches the cyclicality of the
investment dispersion, the cyclicality of the output growth dispersion, and the cyclicality of
the fraction of spike adjusters jointly. This means effectively that the procyclicality of the
investment rate dispersion places a tight bound on the volatility and thus the importance of
firm-level risk shocks. If we increase this volatility and make second moment shocks “more
important,” the procyclicality of the investment rate dispersion disappears.
Why is this important? In the Introduction we mentioned a growing literature that stud-
ies the aggregate consequences of shocks to firm-level productivity risk. To get quantitatively
realistic predictions out of models with such shocks, researchers have to measure the size of
these risk fluctuations. This measurement is invariably assumption-laden. Bachmann and
Bayer (2013) demonstrate that, for instance, depending on whether only continuing firms are
taken into account in the sample, the volatility of firm-level risk can vary substantially. In
contrast, as we show in this paper, the mild procyclicality of the investment rate dispersion
is robust to this issue. Moreover, in order to measure firm-level productivity risk, researchers
have to go through all the steps that need to be taken to calculate investment rates and, in
addition, estimate a production function.
In this paper we use a mid-range estimate for the volatility of firm-level risk: the time-
series coefficient of variation of firm-level risk in our baseline calibration is 4.72%.18 One
approach to dealing with this parameter uncertainty is to report lower and upper bound
scenarios. Our alternative approach is to use more cross-sectional information as an overi-
dentifying restriction and investigate whether the models with risk shocks not only match
the cyclical behavior of the dispersion of the driving force (firm-level revenue productivity
growth) or merely one outcome variable (firm-level sales growth) but jointly the cyclical
behavior of the dispersion of all major firm-level outcome variables. The next section will
illustrate the implications of this alternative approach for the literature on risk shocks in
environments with physical capital adjustment frictions.
18Bloom (2009) uses roughly a four times more volatile shock process for firm-level risk.
21
IV.B Aggregate Fluctuations
How does our baseline model perform in terms of standard aggregate time-series moments?
Table 11 compares unconditional second moments of aggregate output, consumption, invest-
ment and employment between the model in our baseline calibration, a version of the model
without second moment shocks and with frictionless capital adjustment (essentially the RBC
model), and the data. As in the standard RBC model, output volatility is roughly matched,
consumption and employment are somewhat too smooth, and aggregate investment is too
volatile. The model falls short in terms of persistence and overpredicts the comovement of
aggregates with output. Thus we have similar success or a lack thereof as in the standard
RBC model in matching unconditional second moments. Our new focus is on the cross-
sectional dynamics of the model and its aggregate implications, which is why the simple
RBC model, despite its shortcomings, is the right point of departure for our analysis.
Table 11: Aggregate Fluctuations
Y C I N
volatility BL-Model 2.19 0.82 10.22 1.52RBC-Model 2.02 0.76 9.64 1.45Data 2.30 1.79 4.37 1.80
persistence BL-Model 0.29 0.55 0.22 0.20RBC-Model 0.25 0.56 0.18 0.18Data 0.48 0.67 0.42 0.61
correlation with Y BL-Model 1.00 0.87 0.98 0.96RBC-Model 1.00 0.83 0.97 0.96Data 1.00 0.66 0.83 0.68
Notes: The table displays the percent standard deviations (volatility), autocorrelation (persistence), and
correlation with aggregate output of HP(100)-filtered log aggregate output (Y), consumption (C), investment
(I), and employment (N) of the model under the baseline calibration (‘BL-Model’), a version of the baseline
model without second moment shocks and with frictionless capital adjustment (‘RBC-Model’), as well as
German aggregate data from VGR.
More interesting are the implications of firm-level risk shocks in our model. To understand
these implications, we perform two separate exercises. First, we investigate in our baseline
model the effect of an increase in the volatility of firm-level risk – making it “more important”
– on the volatility of aggregate output. We proceed exactly as described for the exercises in
22
Table 10, i.e., we vary σσ over 0, 1, 2, 4. In the extreme case of σσ = 0, firm risk is constant
over time, yet aggregate productivity is still the result of a bivariate stochastic process.
In the second exercise we, counterfactually, eliminate shocks to aggregate productivity in
order to understand the impact on aggregate volatility of risk shocks in isolation. Of course,
this yields not literally a variance decomposition, given the correlated nature of the aggregate
shock process and the nonlinearity of the model, but we believe that we can thus nevertheless
gauge the relevance of idiosyncratic risk shocks for aggregate fluctuations. Again, we vary
σσ over σσ = 0, 1, 2, 4.
Table 12: Aggregate Output Volatilities and the Volatility of Firm-Level Risk Shocks
Full Model Risk-Only Model
Standard Dev. Decline in Standard Dev. Varianceof Output Variance of Output of Output Explained
σσ = 0 2.29 0% - -σσ = 1 (BL) 2.19 9% 0.24 1%σσ = 2 2.12 15% 0.46 4%σσ = 4 2.08 17% 0.82 13%
Notes: The table displays in the left panel (‘Full Model’) the percent standard deviations of HP(100)-
filtered log aggregate output (Y) in our model under the baseline calibration and for σσ = 0, 2, 4. It also
displays the percentage decline in the variance of aggregate output for σσ = 1, 2, 4 relative to σσ = 0:
|var(Y )[σσ=1,2,4]−var(Y )[σσ=0]var(Y )[σσ=0] |. The right panel (‘Risk-Only’) displays the output volatilities for a version of
the baseline model without any aggregate productivity shocks. It also displays the percentage of the variance
of aggregate output in the data that is explained by the ‘Risk-Only Model’-model: var(Y )[σσ=1,2,4]var(Y )[Data] .
Table 12 shows the results from these exercises. We focus on the fluctuations in aggre-
gate output.19 Introducing countercyclical dispersions in idiosyncratic productivity shocks
decreases the variance of output by roughly 9% relative to the case where there are no risk
shocks.20 This change identifies the total business cycle effect of the risk shocks. This effect
includes the real options effect as well as the volatility and the intensive margin effect of
risk. Since risk shocks are countercyclical, a decline in aggregate volatility means that the
volatility effect also dominates in aggregate fluctuations.21 Looking at the right panel of
19The results for the volatility of other aggregates are similar. The comovement patterns and persistenceremain basically unchanged when we vary σσ.
20We focus on the comparison of variances instead of standard deviations, as only the former are mean-ingfully additive.
21There are also so-called Hartman-Abel effects at work here: higher risk concentrates economic activityin highly productive firms and increases aggregate output through Jensen’s inequality.
23
Table 12 we see that the contribution to the output fluctuations in the data for the baseline
calibration of risk shocks is only just above 1%.
If we were to entertain more volatile risk shocks, the effects would get larger. For σσ = 4
the variance of aggregate output would be dampened by over 17%. Risk shocks would
explain 13% of output fluctuations, which in terms of importance would put them in the
neighborhood of conventional monetary policy shocks (see, for instance, Smets and Wouters
(2007)). However, such relatively strong risk fluctuations would be difficult to reconcile with
the procyclicality of the investment rate dispersion, as we have seen in the previous section.
We believe this is a more general insight: the procyclicality of the investment rate dispersion
constrains any model with risk shocks and, thus, a fortiori, any model where risk shocks
have important aggregate implications. This makes the procyclicality of the investment rate
dispersion an important statistic for researchers studying the aggregate effects of risk shocks
to firm-level driving forces.
IV.C Robustness
We have checked the robustness of our quantitative results with respect to a range of modeling
assumptions and calibration choices. The results of these robustness checks are summarized
in Table 13.22 The bottom line is: both the (not targeted) procyclicality of the investment
rate dispersion and the (not targeted) small relevance of risk shocks for aggregate fluctua-
tions, given the procyclicality of the investment rate dispersion, are robust features.
First, we check whether the timing assumption – firms observe today’s productivity
dispersion – is important. When we alternatively assume that firms observe the dispersion
of productivity shocks one period ahead, as some of the literature has done (Bloom (2009),
for instance), we find that the aggregate “importance” of risk shocks increases slightly. But
in terms of the broader economic picture, the results are unchanged.
Second, we check whether the assumption that adjustment costs are in units of labor is
driving the results. In one experiment, we increase the disutilty of labor so that average
time spent at work equals 0.25; without much impact for the baseline results. As a second
experiment, we denote adjustment costs in units of the numeraire. In the baseline spec-
ification, with high wages adjustment is more costly in booms, which somewhat dampens
investment activity: adjustment costs are procyclical in the baseline. In the alternative spec-
ification with acyclical adjustment costs, the procyclicality of the investment rate dispersion
(through the increased procyclicality of the fraction of spike adjusters) increases slightly.
Third, we check whether our results change for deviations from the estimated production
function parameters. Starting from Bloom’s (2009) choice, θ = 0.25, ν = 0.50, as a refer-
22Unless stated otherwise, we recalibrate the adjustment cost parameter to minimize Ψ(ξ) for each robust-ness check.
24
Table 13: Robustness of Quantitative Implications
Full Full Risk OnlyModel Model Model
Cyclicality of Decline in VarianceInvestment Dispersion Variance of Output Explained
Baseline 0.53 9% 1%
σ(εj,t) observed in t− 1 0.43 11% 2%
High disutility of labor 0.35 9% 1%Adj. Costs in numeraire units 0.58 9% 1%
Elasticities θ = 0.25, ν = 0.50 0.56 11% 1%Elasticities θ = 0.27, ν = 0.53 0.46 13% 2%Elasticities θ = 0.28, ν = 0.56 0.33 15% 3%
Frictionless small investment 0.46 9% 1%
High Adj. Costs: ξ = 0.5, σσ = 2 0.30 16% 4%High Adj. Costs: ξ = 0.5, σσ = 4 -0.24 20% 11%
Notes: See notes to Tables 8 and 12. We display |var(Y )[σσ=... ]−var(Y )[σσ=0]var(Y )[σσ=0] | (second column) and
var(Y )[σσ=... ]var(Y )[Data] (third column). Unless stated otherwise, σσ = 1.
ence point23 we increase the returns-to-scale from 0.75 to 0.83 (equivalently, we decrease the
markup from 1.33 to 1.20) and thereby also increase the capital elasticity of static revenue
from 0.50 to 0.63.24 We then recompute the properties of idiosyncratic productivity in the
USTAN data set, imposing the corresponding production function parameters, and recali-
brate the adjustment costs parameter, ξ.25 It is clear that while the cyclicality of investment
23If one views the DRTS assumption in our model as a stand-in for a CRTS production function withmonopolistic competition, then these choices correspond to an employment elasticity of the underlyingproduction function of two-thirds and a markup of 1
θ+ν = 1.33; returns-to-scale, θ + ν, are 0.75, and the
capital elasticity of static revenue, θ1−ν , is 0.5.
24We do so in a way that remains consistent with an underlying CRTS production function with an em-ployment elasticity of two-thirds. Interestingly, with the exception of the two primary sector industriesAgriculture and Mining and Energy, which exhibit a very low procyclicality in their investment rate disper-sion, all other two-digit industries have θ
1−ν lower than 0.63 and θ + ν lower than or equal to 0.83.25Parameters for (θ = 0.25, ν = 0.5): ξ = 0.2, σ(ε) = 0.0919, ρε = 0.9525. Parameters for (θ =
25
dispersion decreases with a production function with less curvature,26 it remains procyclical
and in the ballpark of the baseline calibration. Importantly, a higher volatility of firm-level
risk remains inconsistent with a procyclical investment dispersion under these production
function specifications.
Fourth, our results are robust to an extension of the model that allows for costless small
adjustments of a firm’s capital stock, as proposed by Khan and Thomas (2008), to match
the fraction of small investment rates in the data.
Finally, could larger idiosyncratic risk fluctuations be reconciled with the observed pro-
cyclicality of investment dispersion in the data, if we assume higher adjustment costs, as
Table 9 suggests? The answer is yes, in principle, but quantitatively the evidence still points
toward small risk fluctuations. The last two rows of Table 13 show that raising the adjust-
ment cost parameter from ξ = 0.2 to ξ = 0.5 and at the same time doubling the volatility
of firm risk leads to a correlation of investment dispersion with output of 0.3, compared to
0.45 in the data. Quadrupling the volatility of firm-level risk would lead to countercyclical
investment dispersion (-0.24), as in the baseline calibration. We note that ξ = 0.5 means
that firms pay average adjustment costs per unit of output, conditional on adjustment, in
the upper end of the estimated numbers in the literature: 26% versus 11% in the baseline
(see Bloom (2009), Table IV).
V Final Remarks
The cross-sectional standard deviation of firm-level investment is robustly and significantly
procyclical. This is likely the result of lumpy investment at the micro level and a modest
amount of countercyclical fluctuations in firm-level risk. In an example, namely, the recent
literature on risk shocks in environments with physical adjustment frictions, we show that the
cyclicalities of cross-sectional dispersion measures are jointly informative about the strengths
of these risk shocks. This is a more general insight: the procyclicality of the investment rate
dispersion constrains any model that features shocks to firms’ productivity dispersions. To
the extent that such risk shocks matter for aggregate fluctuations, yet are difficult to measure
directly, the procyclicality of the investment rate dispersion is an informative statistic for
macroeconomics. More generally, our findings suggest that the time-series behavior of the
entire cross-section of firm-level investment is informative on how investment and investment
frictions at the micro level should be modeled. We leave an exploration of this hypothesis
for future research.
0.2667, ν = 0.5333 ): ξ = 0.3, σ(ε) = 0.0914, ρε =0.9375. Parameters for (θ = 0.2778, ν = 0.5556): ξ = 0.4,σ(ε) = 0.0912, ρε = 0.9200.
26Interestingly, across the two-digit industries in USTAN we find a similar pattern between the curvatureof the production function and the procyclicality of investment dispersion.
26
References
Ackerberg, D., Caves, K., and Frazer, G. (2006). Structural identification of production
functions. MPRA Paper 38349, University Library of Munich, Germany.
Arellano, C., Bai, Y., and Kehoe, P. (2012). Financial markets and fluctuations in uncer-
tainty. Federal Reserve Bank of Minneapolis Research Department Staff Report 466.
Bachmann, R. and Bayer, C. (2013). Wait-and-see business cycles? Journal of Monetary
Economics, forthcoming.
Bachmann, R., Caballero, R., and Engel, E. (2013a). Aggregate implications of lumpy invest-
ment: New evidence and a DSGE model. American Economic Journal: Macroeconomics,
forthcoming in October.
Bachmann, R., Elstner, S., and Sims, E. (2013b). Uncertainty and economic activity: Evi-
dence from business survey data. American Economic Journal: Macroeconomics, 5(2):217–
249.
Bayer, C. (2006). Investment dynamics with fixed capital adjustment cost and capital market
imperfections. Journal of Monetary Economics, 53(8):1909 – 1947.
Beaudry, P., Caglayan, M., and Schiantarelli, F. (2001). Monetary instability, the pre-
dictability of prices, and the allocation of investment: An empirical investigation using
U.K. panel data. American Economic Review, 91(3):648–662.
Berger, D. and Vavra, J. (2011). Dynamics of the US price distribution. mimeo, Yale
University.
Bloom, N. (2009). The impact of uncertainty shocks. Econometrica, 77(3):623–685.
Bloom, N., Floetotto, M., Jaimovich, N., S.-E. I., and Terry, S. (2012). Really uncertain
business cycles. NBER WP 18245.
Caballero, R. J. and Engel, E. M. R. A. (1999). Explaining investment dynamics in U.S.
manufacturing: A generalized (S,s) approach. Econometrica, 67(4):783–826.
Caballero, R. J., Engel, E. M. R. A., Haltiwanger, J. C., Woodford, M., and Hall, R. E.
(1995). Plant-level adjustment and aggregate investment dynamics. Brookings Papers on
Economic Activity, 1995(2):1–54.
27
Castaneda, A., Diaz-Gimenez, J., and Rios-Rull, J.-V. (1998). Exploring the income distri-
bution business cycle dynamics. Journal of Monetary Economics, 42(1):93 – 130.
Christiano, L., Motto, R., and Rostagno, M. (2010). Financial factors in economic fluctua-
tions. ECB Working Paper 1192.
Chugh, S. (2012). Firm risk and leverage-based business cycles. mimeo, University of
Maryland.
Cooper, R. W. and Haltiwanger, J. C. (2006). On the nature of capital adjustment costs.
Review of Economic Studies, 73(3):611–633.
Curdia, V. and Reis, R. (2011). Correlated disturbances and U.S. business cycles. mimeo,
Columbia University.
Davis, S., Haltiwanger, J., and Schuh, S. (1996). Job creation and destruction. The MIT
Press.
Davis, S. J., Haltiwanger, J., Jarmin, R., and Miranda, J. (2006). Volatility and dispersion in
business growth rates: Publicly traded versus privately held firms. Working Paper 12354,
National Bureau of Economic Research.
Den Haan, W. J. (2010). Assessing the accuracy of the aggregate law of motion in models
with heterogeneous agents. Journal of Economic Dynamics and Control, 34(1):79–99.
Doepke, J., Funke, M., Holly, S., and Weber, S. (2005). The cross-sectional dynamics of
German business cycles: a bird’s eye view. Discussion Paper Series 1: Economic Studies,
23.
Doepke, J. and Weber, S. (2006). The within-distribution business cycle dynamics of German
firms. Discussion Paper Series 1: Economic Studies, 29.
Doms, M. and Dunne, T. (1998). Capital adjustment patterns in manufacturing plants.
Review of Economic Dynamics, 1(2):409–429.
Eisfeldt, A. and Rampini, A. (2006). Capital reallocation and liquidity. Journal of Monetary
Economics, 53(3):369–399.
Gandhi, A., Navarro, S., and Rivers, D. (2011). On the identification of production functions:
How heterogeneous is productivity? University of Western Ontario, CIBC Centre for
Human Capital and Productivity Working Papers 20119, University of Western Ontario,
CIBC Centre for Human Capital and Productivity.
28
Gilchrist, S., Sim, J., and Zakrajsek, E. (2010). Uncertainty, financial frictions, and invest-
ment dynamics. mimeo, Boston University.
Gourio, F. (2008). Estimating firm-level risk. mimeo, Boston University.
Gourio, F. and Kashyap, A. K. (2007). Investment spikes: New facts and a general equilib-
rium exploration. Journal of Monetary Economics, 54(Supplemen):1–22.
Heckman, J. J. (1976). The common structure of statistical models of truncation, sample
selection and limited dependent variables and a simple estimator for such models. In
Annals of Economic and Social Measurement, Volume 5, number 4, NBER Chapters,
pages 475–492. National Bureau of Economic Research, Inc.
Higson, C., Holly, S., and Kattuman, P. (2002). The cross-sectional dynamics of the US
business cycle: 1950-1999. Journal of Economic Dynamics and Control, 26(9-10):1539–
1555.
Higson, C., Holly, S., Kattuman, P., and Platis, S. (2004). The business cycle, macroeco-
nomic shocks and the cross-section: The growth of UK quoted companies. Economica,
71(282):299–318.
Kehrig, M. (2011). The cyclicality of productivity dispersion. mimeo, Northwestern Univer-
sity.
Khan, A. and Thomas, J. K. (2008). Idiosyncratic shocks and the role of nonconvexities in
plant and aggregate investment dynamics. Econometrica, 76(2):395–436.
Krusell, P. and Smith, A. (1997). Income and wealth heterogeneity, portfolio choice, and
equilibrium asset returns. Macroeconomic dynamics, 1(02):387–422.
Krusell, P. and Smith, Jr, A. (1998). Income and wealth heterogeneity in the macroeconomy.
Journal of Political Economy, 106(5):867–896.
Levinsohn, J. and Petrin, A. (2003). Estimating production functions using inputs to control
for unobservables. Review of Economic Studies, 70(2):317–341.
Narita, F. (2011). Hidden actions, risk-taking and uncertainty shocks. mimeo, University of
Minnesota.
Olley, G. S. and Pakes, A. (1996). The dynamics of productivity in the telecommunications
equipment industry. Econometrica, 64(6):1263–1297.
29
Panousi, V. and Papanikolaou, D. (2012). Investment, idiosyncratic risk, and ownership.
Journal of Finance, 67(3):1113–1148.
Schaal, E. (2011). Uncertainty, productivity and unemployment in the great recession.
mimeo, Princeton University.
Smets, F. and Wouters, R. (2007). Shocks and frictions in U.S. business cycles: A Bayesian
DSGE approach. American Economic Review, 97(3):586–606.
Stoess, E. (2001). Deutsche Bundesbank’s corporate balance sheet statistics and areas of
application. Schmollers Jahrbuch: Zeitschrift fuer Wirtschafts-und Sozialwissenschaften
(Journal of Applied Social Science Studies), 121:131–137.
Tauchen, G. (1986). Finite state markov-chain approximations to univariate and vector
autoregressions. Economics Letters, 20(2):177–181.
Vavra, J. (2012). Inflation dynamics and time-varying uncertainty: New evidence and an Ss
interpretation.
von Kalckreuth, U. (2003). Exploring the role of uncertainty for corporate investment deci-
sions in Germany. Swiss Journal of Economics and Statistics, 139(II):173–206.
Zhang, L. (2005). The value premium. Journal of Finance, LX(1):67–103.
30
A Data
A.1 Description of the Database
Our German firm-level data source is USTAN (Unternehmensbilanzstatistik) of Deutsche
Bundesbank. It provides annual firm-level data from 1971 to 1998 from the balance sheets
and the profit and loss accounts of over 60,000 firms per year. USTAN captures all major
balance-sheet items, the major items of the profit and loss statements, and employment.
Importantly, USTAN provides separate investment data for structures and equipment. As
we will show below, the USTAN sample covers a large fraction of the nonfinancial private
business sector (NFPBS).
It originated as a by-product of the Bundesbank’s rediscounting, i.e. (overnight-)lending
activities. By law, the Bundesbank was required to assess the creditworthiness of all parties
backing a Wechsel, a promissory note or commercial bill of exchange, put up for discounting.
It implemented this regulation by requiring balance-sheet data of all parties involved. These
balance-sheet data were then collected and archived into a database.
Promissory notes were a form of trade credit with widespread use throughout the sample
period. From the volume of a 0.15% stamp tax on promissory notes and bills of exchange, one
can infer that a volume of these titles of roughly 10% of German GDP was issued each year.
Moreover, rediscounting promissory notes was a commonly used instrument of monetary
policy in Germany. Thus, unlike the Federal Reserve, the Bundesbank did not use T-bills
as the major form of collateral but rather private debt. As far as potential cyclical sample
selection is concerned, it is important to note that it had to happen only once in a given
year that a promissory note from a given firm was used as collateral by someone in order for
that firm to appear in USTAN, i.e. it is irrelevant how often that firm issued trade credit
and in what volumes.
The quality of the data is particularly high. All mandatory data collected for USTAN
have been double-checked by Bundesbank staff. The Bundesbank itself frequently uses the
USTAN data for its macroeconomic analyses and for cross-checking national accounting data.
We take this as an indication that the bank considers the data as sufficiently representative
and of high quality.
One drawback of USTAN is that with the introduction of the euro, the Bundesbank
stopped buying commercial bills and collected firm balance-sheet data only irregularly and
only from publicly available sources. For this reason, the data set stops being useful in
1999. Therefore, we only use data from 1971 to 1998, which leaves us, after lagging and
first-differencing, with 26 years of observations from 1973 to 1998.
The coverage of the sample is broad, although it is technically not a representative
31
sample due to the nonrandom sample design. It was also more common to use promissory
notes as trade credit in manufacturing and for incorporated companies, which biases our
data somewhat toward these kinds of firms. And, of course, the Bundesbank would only
rediscount notes to which it gave a good rating, so that the set of firms in USTAN is also
somewhat biased toward financially healthy and larger firms. Nevertheless, USTAN covers
a wide range of firms, in fact a wider range in some dimensions (size, ownership, industry)
than comparable U.S. data sets (ASM, COMPUSTAT), since short-term financing through
promissory notes was a common practice for many German companies across most business
sectors. Due to the Bundesbank’s rediscounting policy, bills of exchange were very liquid for
the creditor.
A.2 Sample Selection
We start with the universe of observations in the USTAN data, merging the files for 1971-
1986 and 1987-1998. In a first pass, we then drop all balance sheets that are irregular, e.g.,
bankruptcy or closing balance sheets, or that stem from a group/holding (Konzernbilanz ).
This leaves us with only regular balance sheets (Handelsbilanz or Steuerbilanz ). We also
drop all firms with missing payroll data or missing or negative sales data, which are basically
nonoperating firms. A small amount of duplicate balance sheets is removed as well. Finally,
we drop the following sectors: hospitality (hotels and restaurants), which only has a small
number of firms in the database, financial and insurance institutions, the mostly public
health and education sectors, as well as other public companies like museums, etc., and
some other small service industries, such as hair cutters, dry cleaners and funeral homes;27
or when sectoral information was missing. The sectoral aggregate we are studying can be
roughly characterized as the nonfinancial private business sector in Germany. This sample
selection leaves us with an initial data set of 1,764,846 firm-year observations and 259,614
different firms. The average number of firms per year is 63,030.
From this initial sample we remove step-by-step observations, in order to get an economi-
cally meaningful data set. We first drop observations from likely East German firms to avoid
a break in the series in 1990. We identify a West German firm as a firm that has a West
German address or has no address information but enters the sample before 1990. Then we
recompute capital stocks with a perpetual inventory method (PIM). In the PIM we drop a
small amount of outliers. We remove observations that do not have a log value-added and a
log capital stock after PIM.
Another part of the data is removed when firms did not have changes in log firm-level
27The number of firms from the public sector and these small industries is tiny to begin with, as they didnot regularly use bills of exchange as a financing instrument.
32
employment (N), capital (K) and real value-added (VA), which obviously requires us to
observe firms for two consecutive years. Then we remove outliers in factor changes and real
value-added changes. Specifically, we identify as outliers in our sample a firm-year in which
the firm-level investment rate or log changes in firm-level real value-added, employment and
capital stock fall outside a three-standard-deviations band around the firm and sectoral-year
mean. Then we compute firm-level Solow residuals and similarly remove observations with
missing log changes in Solow residuals as well as outliers therein. We finally remove – before
and after each step of the outlier removal – firms that have less than five observations in
firm-level Solow residual changes. We conduct extensive robustness checks of our results
to the choices for the outlier and observation thresholds. Table 14 summarizes how many
observations are dropped in each step.
Table 14: Sample Creation
Criterion Firm-Year ObservationsInitial Sample 1,764,846East Germany -104,299Outliers in PIM -7,539Missing log value-added -1,349Missing log capital -31,819Missing log-changes in N, K, VA -161,668Outliers in N, K and VA log-changes -41,453Missing log-changes in Solow residual -126,086Outliers in Solow residual log-changes -18,978Not enough observations -417,550Final Sample 854,105
A.3 Sample Composition
The final sample then consists of 854,105 firm-year observations, which amounts to observa-
tions on 72,853 different firms. The average observation length of a firm in the sample is 11.7
years. The average number of firms per year is 32,850. The following Tables 15, 16 and 17
show the average industry,28 the legal form and the size distributions in our final sample.
USTAN’s industry coverage, while somewhat biased toward manufacturing firms, includes
the construction, service and the primary sectors. While a bias toward larger firms remains,
the size coverage is still fairly broad: 31% of all firm-year observations in our final baseline
28WZ 2003 is the industry classification from 2003 that the German national accounting system (Volk-swirtschaftliche Gesamtrechnung, VGR) uses.
33
sample have fewer than 20 employees and 57% have fewer than 50 employees. In terms of
ownership structure, only 2% of firm-year observations are from publicly traded firms, just
under 60% from limited liability companies and just under 40% from private firms with fully
liable partners.
Table 15: Two-Digit Industry Distribution
ID Sector Observations Frequency WZ 200310 Agriculture 12,291 1.44% A, B20 Energy & Mining 4,165 0.49% C, E31 Chemical Industry, Oil 14,721 1.72% DF, DG32 Plastics, Rubber 23,892 2.80% DH33 Glass, Ceramics 28,623 3.35% DI34 Metals 30,591 3.58% DJ35 Machinery 162,407 19.01% DK, DL, DM, DN36 Wood, Paper, Printing 61,672 7.22% DD, DE37 Textiles, Leather 46,173 5.41% DB, DC38 Food, Tobacco 37,708 4.41% DA40 Construction 54,569 6.39% F61 Wholesale Trade 213,071 24.95% G5162 Retail Trade & Cars 142,137 16.64% G50, G5170 Transportation & Communication 22,085 2.59% I
Total 854,105 100%
Table 16: Legal Form Distribution
Legal Form Observations FrequencyPublicly Traded (AG, KGaA, etc.) 18,582 2.18%Limited Liability Companies (GmbH, GmbH&Co., etc.) 506,184 59.26%Fully Liable Partnerships (OHG, KG, etc.) 327,526 38.35%Other: unincorporated associations (e.V.) 1,813 0.21%municipal agencies (Korperschaften oR) etc.Total 854,105 100%
34
Tab
le17
:SizeDistributionsofFirms
Numberof
Employees
1-4
5-9
10-1
415
-19
20-4
950
-99
100-
249
250-
499
500+
Fra
ctio
n6.
14%
9.46
%8.
24%
7.30
%26
.28%
17.0
4%14
.37%
5.68
%5.
49%
CapitalSto
ck(i
n10
0019
91-E
uro
)0-
299
300-
599
600-
999
1,00
0-1,
499
1,50
0-2,
499
2,50
0-4,
999
5,00
0-9,
999
10,0
00-2
4,99
925
,000
+F
ract
ion
8.23
%9.
01%
9.67
%9.
36%
13.0
8%17
.71%
13.8
7%11
.08%
7.99
%
RealValueAdded
(in
1000
1991
-Euro
)0-
299
300-
499
500-
999
1,00
0-1,
499
1,50
0-2,
499
2,50
0-4,
999
5,00
0-9,
999
10,0
00-2
4,99
925
,000
+F
ract
ion
8.17
%7.
93%
16.3
8%11
.56%
14.4
5%16
.28%
11.2
0%8.
25%
5.79
%
35
A.4 Perpetual Inventory Method
In order to obtain economically meaningful stocks of capital series for each firm, we have
to re-calculate capital stocks in a Perpetual Inventory Method (PIM); see Bayer (2006), for
instance. The first step is to compute firm-level investment series, Ij,t, from the corporate
balance sheets, which contain data only on accounting capital stocks, kaj,t, and accounting
total depreciation, daj,t. The following accumulation identity for the book value of capital
allows us to back out nominal firm-level investment, pIt Ij,t:29
kaj,t+1 = kaj,t − daj,t + pIt Ij,t. (10)
The next step is to recognize that capital stocks from corporate balance sheets are not
directly usable for economic analysis for two reasons: 1) accounting depreciation, daj,t, in
corporate balance sheets is often motivated by tax reasons and is typically higher than
economic depreciation, δej,t, expressed as a rate; 2) accounting capital stocks are reported at
historical prices. Both effects would lead to an underestimation of the real firm-level capital
stock, if one were to simply deflate the current accounting capital stock, kaj,t, with a current
investment price deflator, pIt (assuming that pIt increases over time). We therefore apply a
Perpetual Inventory Method (PIM) to compute economic real capital stocks:
k(1)j,1 = kaj,1. (11)
k(1)j,t+1 = (1− δet ) k
(1)j,t +
pItpI1991,t
Ij,t. (12)
kaj,1 is the accounting capital stock in 1991 prices at the beginning of an uninterrupted
sequence of firm observations – if for a firm-year we have a missing investment observation,
the PIM is started anew when the firm appears again in the data set. The investment-good-
price deflator is pI1991,t, with 1991 as the base year. We estimate the economic depreciation
rate δet for each year from national accounting data, VGR, separately for equipment and
nonresidential structures (Table 3.1.3, VGR, Nettoanlagevermogen nach Vermogensarten in
jeweiligen Preisen, Ausrustungen und Nichtwohnbauten; Table 3.1.4, VGR, Abschreibungen
nach Vermogensarten in jeweiligen Preisen, Ausrustungen und Nichtwohnbauten). VGR
contains sectoral and capital-good-specific depreciation data only after 1991, which is why
29Specifically, kaj,t is the sum of balance-sheet items ap65, Technische Anlagen und Maschinen, andap66, Andere Anlagen, Betriebs-und Geschaftsausstattung, for equipment; and balance-sheet item ap64,Grundstucke, Bauten, for structures. Since balance-sheet data are typically end-of-year stock data, noticethat kaj,t is the end-of-period capital stock in year t− 1. daj,t is profit and loss account item ap156, Abschrei-bungen auf Sachanlagen und immaterielle Vermogensgegenstande des Anlagevermogens. In contrast to kaj,t,daj,t is not given for each capital good separately. For the solution of this complication, see below.
36
we decided to use only capital-good-specific depreciation rates for the entire time horizon.
For the data sources for investment price deflators, see footnote 32 below. The drawback
to this procedure is that we do not directly observe capital-good specific daj,t in the balance
sheets, so that (10) is not directly applicable to the two types of capital good separately. We
therefore split up daj,t according to the fraction that each type of capital good accounts for in
the book value of total capital, weighting each type of capital good by its VGR depreciation
rate. We finally aggregate both types of capital into a single capital good at the firm level.
There is a final complication, which arises through relying on kaj,1 as the starting value
of the PIM. The recalculation of capital stocks is motivated by the bias that historical cost
accounting and tax depreciation induce, i.e., that the book value of capital is typically not a
good estimate of the productive real capital stock of the firm at that time. To take this issue
into account also for the first observation of a firm, we calculate the time-average factor φ
(for each sector), by which k(1)j,t is larger than kaj,t, and replace kaj,1 by φkaj,1 in the perpetual
inventory method. We do this iteratively until φ converges, i.e., we calculate (using k(0)j,t = kaj,t
and φ(0) = 1):
k(n)j,t+1 = (1− δet ) k
(n)j,t +
pItpI1991,t
Ij,t (13)
k(n)j,1 = φ(n−1)k
(n−1)j,1 (14)
φ(n) = (NT )−1∑j,t
k(n)j,t
k(n−1)j,t
(15)
We stop when for each sector and each capital good category φ < 1.1.30
Since we want to compute economic, i.e. productive, capital stocks, we then – as a final
step – add to the capital stock series from this iterative PIM the net present value of the
real expenditures for renting and leasing equipment and structures.31
30Extreme φ’s indicate that constant depreciation is not a good approximation for this particular firm.Such a firm will have had an episode of extraordinary depreciation (e.g., fire, accident, etc.) and the capitalstocks by PIM will be a bad measure of the actual capital stock after the accident. That is why we drop asmall number of observations from the top and bottom of the φ-distribution (see Table 14).
31 We take item ap161, Miet- und Pachtaufwendungen, from the profit and loss accounts, deflate it by theimplicit investment good price deflator, which we compute from Tables 3.2.8.1 and 3.2.9.1 from VGR, andthen divide it by a measure of the user cost of capital. The latter is simply the sum of real interest rates fora given year, which we compute from nominal interest rates on corporate bonds and ex-post CPI inflationdata, and the time-average, accounting capital-good-weighted depreciation rate per firm.
37
A.5 Representativeness
How well does the USTAN aggregate represent the nonfinancial private business sector (NF-
PBS) in Germany? Table 18 shows that USTAN represents on average 70% of the value-
added of the NFPBS, 44% of its investment, etc.32
Table 18: USTAN and the NFPBS
USTAN/NFPBSValue-Added 70%Investment 44%Capital 71%Employment 49%Payroll 54%
Table 19 shows that the cross-sectional averages of investment as well as output, employ-
ment and productivity growth, computed from USTAN, are strongly positively correlated
with the cyclical component of the real gross value-added of the nonfinancial private busi-
ness sector. This means that USTAN represents well the cyclical behavior of the sectoral
aggregate it is meant to represent.
32 NFPBS value-added is taken from Bruttowertschopfung in jeweiligen Preisen, Table 3.2.1 of VGR,deflated by the implicit deflator for aggregate value-added, Table 3.1.1 of VGR (we apply the same deflatorto the USTAN data). The base year is always 1991. NFPBS investment is Bruttoanlageinvestitionen injeweiligen Preisen from Table 3.2.8.1, deflated with the implicit sector-specific investment price deflatorsgiven by Bruttoanlageinvestitionen - preisbereinigt, a chain index, from Table 3.2.9.1, VGR. NFPBS capital isNettoanlagevermogen in Preisen von 2000 from Table 3.2.19.1, VGR, re-chained to 1991 prices. In computingboth the investment and the capital data for USTAN in the PIM, we use the implicit sector and capital-good-specific (equipment and nonresidential structures) deflators for investment: Tables 3.2.8.2, 3.2.9.2.,3.2.8.3 and 3.2.9.3., VGR. We also experiment with deflating USTAN data with a uniform investmentprice deflator, the Preisindex der Investitionsguterproduzenten, source: GP-X002, Statistisches Bundesamt.NFPBS employment is the number of employed, Arbeitnehmer, from Table 3.2.13, VGR. Finally, payrollis taken from Arbeitnehmerentgelt, Table 3.2.10., VGR, deflated by the same general implicit deflator foraggregate value-added that we use to deflate value-added numbers.
38
Table 19: Cyclicality of Cross-Sectional Averages
Cross-sectional Moment Correlation with Cyclemean(ij,t) 0.756∗∗∗
mean(∆ log yj,t) 0.663∗∗∗
mean(∆nj,t
0.5∗(nj,t−1+nj,t)) 0.602∗∗∗
mean(∆ log εj,t) 0.592∗∗∗
Notes: The table shows the correlation with the cycle of the cross-sectional averages, linearly detrended,
of, respectively, the investment rate, the log-change of real gross value-added (we deflate the profit and
loss account item ap153, Rohergebnis, with the aggregate value-added deflator from VGR data), the net
employment change rate, and the log-change of Solow residuals, all at the firm level. We have removed firm
fixed and 2-digit industry-year effects from each variable. As a cyclical indicator we use the HP(100)-filtered
aggregate real gross value-added in the German nonfinancial private business sector, computed from German
VGR (Volkswirtschaftliche Gesamtrechnungen) data. ∗∗∗ indicates significance at the 1%level, resulting from
an overlapping block bootstrap of four-year windows with 10,000 replications.
A.6 The Cross-sectional Investment Rate Distribution
As Table 20 shows, the distribution of firm-level investment rates from our USTAN sample
is comparable to the one calculated for the U.S. from the LRD, reported in Cooper and
Haltiwanger (2006). For comparability with Cooper and Haltiwanger (2006), we also show
the distribution of investment rates forIj,tkj,t
. The USTAN sample exhibits less disinvestment
activity, slightly more inactivity, and a little less frequent spikes. There are various reasons
for these differences. First, LRD is plant-level data, whereas our data is firm level, so some
difference may come from unit aggregation, in particular, observing fewer spikes in USTAN.
Second, the U.S. and German manufacturing sector were facing fairly different compositional
trends over the respective sample periods, which may explain why more negative investment
due to stronger reallocation is observed in the LRD.
39
Table 20: Investment Rate Distributions in USTAN
Negative Inactivity Positivespike intermediate intermediate spike
ij,t
All firms (USTAN) 0.4% 2.6% 14.8% 68.9% 13.4%Manufacturing (USTAN) 0.4% 2.0% 11.1% 74.7% 11.9%
Ij,tkj,t
All firms (USTAN) 0.3% 2.6% 15.1% 67.7% 14.2%Manufacturing (USTAN) 0.3% 2.0% 11.4% 73.6% 12.7%
LRD 1.8% 8.6% 8.1% 62.9% 18.6 %
Notes: ij,t =Ij,t
0.5(kj,t+kj,t+1)denotes our baseline investment rate definition.
Ij,tkj,t
denotes the definition of
the investment rate used by Cooper and Haltiwanger (2006). An investment spike is defined, for either
investment rate, as being larger than 20% in absolute value. Inactivity is defined as an investment rate that
is smaller than 1% in absolute value. The LRD data are taken from Cooper and Haltiwanger (2006).
A.7 Estimating the Production Function
We estimate the coefficients θ, ν of the production function by the median of the firm average
share of factor expenditure in total value-added, as defined by:33
νj = T−1j
∑t
wj,tnj,tyj,t
θj = T−1j
∑t
(rt + δj)kj,tyj,t
In a frictionless setup, this is estimating the production function coefficients from the
first-order conditions. Importantly, this estimator is robust to classical measurement error
in capital and labor. We take as the real interest rate, rt, the average return on corporate
bonds minus the ex post inflation rate and calculate firm-specific depreciation rates, δj, from
capital-good-specific VGR depreciation rates, weighted by the firm-specific capital good
portfolio.
33We use profit and loss account item ap153, Rohergebnis, for firm-level value-added and profit and lossaccount item ap154, Personalaufwand, for the firm-level wage bill.
40
Under the null hypothesis of our model, i.e., nonconvex capital adjustment frictions, these
first-order conditions do not hold exactly for capital. However, when we estimate the pro-
duction function coefficients from data simulated from our model in data sets of comparable
size to USTAN, we find that this simple estimation procedure from factor expenditure shares
works remarkably well. In fact, the labor share is estimated exactly, as labor can be adjusted
without frictions in the model. The capital share estimated from the model-simulated data
is θ = 0.2238, with the true θ being 0.2075.34
Alternative estimation approaches, such as those advocated by Olley and Pakes (1996)
or Levinsohn and Petrin (2003), suffer from the collinearity issues discussed in Ackerberg
et al. (2006) and Gandhi et al. (2011). In fact, when we estimate the production function
parameters from model-simulated data using Olley and Pakes’ (1996) estimator, we obtain a
greatly upwardly biased estimate for ν: νOP = 0.9962. The estimated coefficient for capital
is virtually zero. Also for the USTAN data we obtain fairly high coefficients on labor and
low coefficients on capital. For manufacturing, for instance, we obtained Olley and Pakes
estimates of ν = .744 and θ = .069, which is another indication that our model is a good
description of firm-level behavior in the USTAN data set. The Olley and Pakes estimates
are essentially unaffected by the inclusion (or lack thereof) of a selection term (cf. Appendix
B) or the order of approximation used in the third stage of the estimator. In the first stage,
we use only observations where the absolute value of the investment rate exceeds 20%.35
In summary, under the null hypothesis of our model, the factor shares are good estimators
of the production function parameters despite the capital adjustment frictions.
A.8 The Idiosyncratic Productivity Process
We estimate the log-productivity residuals implied by the estimates for the production func-
tion coefficients as:
εj,t = log(yj,t)− zt − ν log nj,t − θ log kj,t.
We then specify for εj,t
εj,t = εj,t + xj,t + µj
εj,t = ρεj,t−1 + uj,t,
34We thus find a small upward bias for the capital coefficient θ. Yet, an exact calibration alongside theadjustment cost parameter using our model would be prohibitively time-consuming and, given the negligiblesize of the bias, would not change our results substantively.
35Using a 10% threshold yields basically the same results.
41
where xj,t denotes measurement error. In order to estimate ρ we allow the measurement error
to be a second- or third-order moving average process. This yields an estimation equation
εj,t = ρεj,t−1 +∑s=1...J
ξs∆εj,t−s + (1− ρ)µj + ζj,t,
where we experiment with J = 3 and J = 4. An unbiased estimate of ρ can be obtained
by an instrumental variable regression, where one uses lagged differences of ∆εj,t−J−1 as
instruments for εj,t−1. The estimated ρ is 0.960 and 0.974 for J = 3 and J = 4, respectively.
The coefficient used in our calibration is 0.9675 - the midpoint of these two estimates.
When measuring the size and the cyclicality of the dispersions of productivity growth, we
apply a slightly simplified approach, given the relatively high values of ρ (and low values of ξs)
that we estimate. Since a specification with a long moving average term for the measurement
error would force us to discard many aggregate data points for generating the necessary lags,
we specify εj,t as a random walk cum classical measurement error, for estimating the variance
of the shocks as36
E∆ε2j,t − σ2me,
where the variance of the measurement error, ∆xj,t, is estimated by the sample analogue to
σ2me = −E(εj,t − εj,t−2)2 + 2E∆ε2j,t
= −E(εj,t − εj,t−2 + xj,t − xj,t−2)2 + 2E(∆εj,t + ∆xj,t)2
= −E(∆εj,t + ∆εj,t−1 + xj,t − xj,t−2)2 + 2E(∆εj,t + ∆xj,t)2
= −E(−σ∆ε,t − σ∆ε,t−1 − 2σx + 2σ∆ε,t + 4σx) = 2σx.
B Robustness of the Empirical Findings
This appendix provides the details for the robustness checks discussed in Section I.C. Ta-
ble 21 shows robustness with respect to the cyclical indicator. We experiment with differ-
ent HP-filter smoothing parameters, with dynamic correlations, with excluding the post-
reunification period and with excluding the years with the most extreme investment disper-
sion observations.37
Table 22 provides robustness checks with respect to sample composition. The first two
36We have conducted a Monte Carlo analysis that shows that the mistake one makes when our measurementerror estimation is applied to not-quite-unit-root data is small. We therefore prefer the simple estimationprocedure for the idiosyncratic shock variance.
37For the sake of brevity, we do not show results for firm-level Solow residual growth and fractions ofadjusters. Results for both are robust: for the former, similar to the values for output growth, for the latter,similar to the values for the fraction of spike adjusters.
42
rows (dropping the first three observations per firm or looking only at firms that are virtually
always in the sample) show that our results are neither driven by firm entry (into the sample)
nor by firm exit, nor do our results depend on how we remove outliers from the sample.
Further robustness checks for the procyclicality of the investment rate dispersion are
available in Table 23. We check for the robustness with respect to replacing the standard
deviation with the interquartile range as the measure of dispersion, and to alternative ways
of detrending the dispersion series. Perhaps most important, the last row of Table 23 shows
that the procyclicality of the investment rate dispersion is not due to cyclical variations
in the sample composition. In the scenario ‘Selection correction’ we control for sample
selection in the following way: we estimate a simple selection model, where lagged firm-
level Solow residuals determine selection and the firm-level investment rate is modeled as
a mean regression. We use the maximum likelihood estimator by Heckman (1976) to infer
the selection-corrected variance of the residual in the firm-level investment rate equation.
The latter is very close to the sample variance of firm-level investment rates, indicating that
our results are not influenced by systematic sample drop outs. While the first stage of the
regression shows that there is a positive selection in terms of levels (more productive firms
being more likely to be in the sample), there is no strong selection with respect to changes
or investment rates and their dispersions.
Finally, Table 24 shows that the ownership structure matters for cross-sectional results
(focusing on publicly traded firms in Germany would eliminate the procyclicality of invest-
ment dispersion), making it important to use broader data sets for the study of cross-sectional
facts (see Davis et al. (2006), for a similar point). Figure 1 plots the time-series of the cross-
sectional standard deviation of firm-level investment rates, linearly detrended and detrended
with an HP(100)-filter, the time-series of the cross-sectional interquartile range of firm-level
investment rates, linearly detrended, and the fraction of spike adjusters, linearly detrended,
against the cyclical component of aggregate real gross value-added.
43
Table 21: Robustness Checks I – Cyclical Indicator
Correlation of ... with
Cyclical Indicator / Sample std(ij,t) std(∆yj,t) std(∆nj,t
0.5∗(nj,t−1+nj,t)) Fraction of
Spike AdjustersBaseline: HP(100)-filtered 0.45∗∗ −0.45∗ −0.50∗∗ 0.61∗∗∗
Real Gross Value-Added, Y
HP(6.25)-filtered Y 0.37∗∗ −0.47∗∗ −0.53∗∗ 0.44∗∗∗
HP(100)-filtered I 0.72∗∗∗ −0.30 −0.31 0.78∗∗∗
Lag of HP(100)-filtered Y 0.26 0.12 0.04 0.36∗
Lead of HP(100)-filtered Y 0.35 −0.56∗∗ −0.54∗∗ 0.48∗∗
Pre-reunification: 1973-1990 0.30 −0.65∗∗∗ −0.63∗∗∗ 0.51∗∗
Exclude min and max std(ij,t) years 0.54∗∗∗ −0.49∗∗ −0.53∗∗ 0.66∗∗∗
Exclude two max |std(ij,t)| years 0.50∗∗∗ −0.58∗∗∗ −0.57∗∗ 0.63∗∗∗
Notes: This table shows the correlation with the cycle of the cross-sectional standard deviations, linearly
detrended, of, respectively, the investment rate, the log-change of real gross value-added and the net employ-
ment change rate, all at the firm level. It also shows, in the last column, the correlation with the cycle of the
average fraction, linearly detrended, of firms exhibiting an investment spike. All data are from the Bundes-
bank’s USTAN database. We have removed firm fixed and 2-digit industry-year effects from each variable.
As a cyclical indicator we use aggregate real gross value-added and aggregate investment in the German
nonfinancial private business sector, computed from German VGR (Volkswirtschaftliche Gesamtrechnun-
gen) data. ∗∗∗,∗∗ ,∗ indicate significance at the 1%, 5%, and the 10% level, respectively, resulting from an
overlapping block bootstrap of four-year windows with 10,000 replications.
Table 22: Robustness Checks II – Sample Composition
Correlation of ... with HP(100)-filtered Y
Sample std(ij,t) std(∆yj,t) std(∆nj,t
0.5∗(nj,t−1+nj,t)) Fraction of
Spike AdjustersW/o entry: drop first 3 obs. per firm 0.39∗ −0.48∗ −0.52∗ 0.61∗∗∗
Only firms with 20+ obs. 0.39∗∗∗ −0.38∗∗ −0.37∗∗ 0.77∗∗∗
Stricter outlier removal (2.5 std.) 0.45∗∗ −0.45∗ −0.55∗∗ 0.62∗∗∗
Looser outlier removal (5 std.) 0.42∗∗ −0.44∗ −0.22 0.62∗∗∗
Percentile outlier removal (5%) 0.59∗∗∗ −0.45∗∗ −0.60∗∗∗ 0.64∗∗∗
Percentile outlier removal (1%) 0.48∗∗ −0.47∗∗ −0.42∗ 0.62∗∗∗
Notes: See notes to Table 21.
44
Table 23: Robustness Checks III – Other
Correlation of investment dispersion with HP(100)-filtered Y
Baseline 0.45 **
iqr(ij,t) 0.57 ***Raw data - no fixed effects 0.45 ***Uniform price index for investment 0.43 **
std(Ij,tkj,t
) (dropping 5% outliers) 0.60 ***
std(Ij,tkj,t
) (dropping 1% outliers) 0.48 **
std(ij,t)quadratic detrending 0.56 ***std(ij,t)cubic detrending 0.60 ***std(ij,t)HP(100)-detrending 0.62 ***
Outlier ≥ 3 std means merger 0.42 **Shorter in sample (2 obs.) 0.44 **Selection correction 0.38 **
Notes: See notes to Table 21. iqr(ij,t) refers to the cross-sectional interquartile range of firm-level investment
rates. ‘Raw data - no fixed effects’ uses the standard deviation of the raw firm-level investment rates, no
fixed effects removed. ‘std(Ij,tkj,t
) (dropping 5% outliers)’ usesIj,tKj,t
the definition of the investment rate used
by Cooper and Haltiwanger (2006). To take care of the higher sensitivity to outliers, we use a 5% outlier
criterion here. ‘std(Ij,tkj,t
) (dropping 1% outliers)’ does the same with a 1% outlier criterion. ‘Uniform price
index for investment’ refers to a scenario, in which we deflate firm-level investment and capital with an
aggregate price deflator for investment goods, not with one-digit industry- and capital-good-specific ones.
The next three rows show the results, when we detrend std(ij,t) not with a linear trend, but, respectively,
with a quadratic, cubic trend and an HP(100)-filter. ‘Outlier ≥ 3 std means merger’ refers to a scenario, in
which we treat an observation of 3 standard deviations above or below the year-specific mean as indicating a
merger and mark the firm henceforth as a new one. ‘Shorter in sample (2 obs.)’ refers to a scenario, in which
we require firms to have two observations in first differences (instead of five) to be in the sample. ‘Selection
correction’ refers to a scenario where we estimate a simple selection model, where lagged firm-level Solow
residuals determine selection and the firm-level investment rate is modeled as a mean regression. We use the
maximum likelihood estimator by Heckman (1976) to infer the selection-corrected variance of the residual
in the firm-level investment rate equation.
45
Table 24: Cyclicality of Cross-Sectional Investment Dispersion – Legal Form
Aggregate Publicly Traded Limited Liability Companies Fully Liable Partnerships
0.45** 0.10 0.32* 0.64***
Notes: See notes to Table 21. ‘Publicly Traded’ means the German legal forms of AG and KGaA. ‘Lim-
ited Liability Companies’ means the German legal forms of GmbH and GmbH & Co KG. ‘Fully Liable
Partnerships’ means the German legal forms of GBR, OHG and KG.
Figure 1: Cyclicality of Cross-sectional Moments - Data
1975 1980 1985 1990 1995−3
−2
−1
0
1
2
Investment std − Lin. Det.Cyclical Component of GDP
1975 1980 1985 1990 1995−3
−2
−1
0
1
2
3
Investment iqrCyclical Component of GDP
1975 1980 1985 1990 1995−3
−2
−1
0
1
2
3
Investment std − HPCyclical Component of GDP
1975 1980 1985 1990 1995−3
−2
−1
0
1
2
3
Fraction of Spike AdjustersCyclical Component of GDP
Notes: See notes to Table 21. For better readability, in each panel we normalize the cyclical components of
the cross-sectional moments and of aggregate real gross value-added by their respective standard deviations.
46
C Numerical Implementation
C.1 Details
C.1.1 Decision Problem
The computable dynamic programming problem of a firm in our model has a 6-dimensional
state space:(k, k)
are the endogenous states, while(ε, z, s, std(ε)
)are the exogenous states.
Since the employment problem has an analytical solution, there is essentially just one con-
tinuous control, k′. We discretize the state space as follows:
1. k: nk = 41 grid points from [0, 40], with a smaller grid width at low capital levels,
where the curvature of the value function is higher.
2. k : nk = 4 grid points: [0.8, 1.0, 1.2, 1.4].
3. ε: nε = 19. The grid points are equi-spaced (in logs) and the width between the
midpoint, which is normalized to unity, and the extreme points is given by 3×√
σ(ε)2
1−ρ2ε,
i.e., three times the unconditional variance of idiosyncratic productivity.
4. z : nz = 5 grid points: [0.9561, 0.9778, 1.0000, 1.0227, 1.0459].
5. s : ns = 5 grid points: [0.0764, 0.0834, 0.0905, 0.0976, 0.1047].
6. std(ε) : nstd(ε) = 3 grid points: [0.30, 0.35, 0.40].
The various transition matrices for the stochastic processes are calculated as follows:
first, using a bivariate version of the procedure in Tauchen (1986), we compute a discrete
bivariate Markov chain on z× s, using the results in equation (9). Second, we then compute
(in the baseline case) for each s a transition matrix on the fixed (across s) ε-grid. The
transition matrix takes two features into account: time-varying σ(ε) and the (small) excess
kurtosis of the idiosyncratic productivity process in the data: 4.4480 on average. Since in
our calibration strategy the fixed adjustment costs parameter is identified by the kurtosis
of the firm-level investment rate (together with its skewness), we want to avoid attributing
excess kurtosis in the firm-level investment rate to lumpy investment, when the idiosyncratic
driving force itself has excess kurtosis. We incorporate the measured excess kurtosis into
the discretization process for the idiosyncratic productivity state by using a mixture of
two Gaussian distributions: N(0, 0.0586) and N(0, 0.1224) - the standard deviations are
0.0905± 0.0319, with a weight of 0.4118 on the first distribution.
Since we allow for a continuous control, k, and k can take on any value continuously,
we can only compute the value function exactly at the grid points above and interpolate for
47
in-between values. This is done by using a multidimensional cubic splines procedure, with a
so-called “not-a-knot” condition, to address the degrees-of-freedom problem that arises when
splines are used. We compute the solution by value function iteration, using 10 steps of policy
improvement after each actual optimization step. The optimum is found by using a golden
section search. Upon convergence, we check single-peakedness of the objective function, to
guarantee that the golden section search is reasonable. We have also experimented with finer
grids for the baseline case and found our results to be robust.
C.1.2 Equilibrium Simulation
For the equilibrium simulations we draw one random series for the aggregate states and fix
it across models. We use T = 1600 and discard the first 500 observations, when we compute
statistics from these simulations.
As in the firm’s decision problem, we use a golden section search to find the optimal
target capital level, given p, at every point in time during the simulation. We find the
market-clearing intertemporal price, using a combination of bisection, secant and inverse
quadratic interpolation methods. The precision of the market-clearing algorithm is better
than 10−7 at every point in time during the simulation.
There is a final complication due to the nature of the bivariate aggregate shock process:
given the correlated processes for aggregate productivity and idiosyncratic firm-level risk,
not all of the 25 (5 × 5) distinct aggregate state combinations are reached with sufficient
frequency during our T = 1600-simulations to compute the regressions (8a) and (8b) state
by state. Since going much beyond T = 1600 would be prohibitively burdensome in terms of
computational resources and time, we proceed as follows: if we have at least five observations
on an aggregate state combination, we run the state-by-state regressions. Otherwise, we use
a version of (8a) and (8b) where we treat z and σ(ε) as if they were continuous variables,
include several, but not all possible interaction terms, and run OLS regressions on these
modified rules to extrapolate the coefficients for the remaining aggregate state combinations.
As we will show below, these somewhat restricted regressions nevertheless provide a good fit
for the time-series of aggregate capital and the marginal utility of consumption.
C.2 Quality of Numerical Approximations
What role does log std(ε) play in the Krusell and Smith (1998) or KS rules, (8a) and (8b)?
Table 25 shows that both our cross-sectional results and, to some extent, our aggregate
results depend crucially on the addition of log std(ε) in the KS rules. For instance, the
procyclicality of the investment rate dispersion is substantially lower without it.
48
Tables 26, 27 and 28 illustrate why. First, the R2 is substantially higher and the standard
error of estimation is substantially lower when log std(ε) is added in the KS rules, especially
the KS rules for the aggregate capital stock. Second, as Den Haan (2010) argues, it is
important to check the quality not only of the one-step-ahead forecasts of the KS rules but
also of forecasts at longer horizons. The performance of the KS rules – measured by the
mean squared percentage deviation of applying the forecasting rules t times (using the actual
realizations of the driving processes) from the actual value on the equilibrium simulation
path (assuming households use the converged one-step-ahead forecasting rules), measured
by the maximum absolute percentage deviation and measured by the correlation coefficient
between the t-forecast and the actual value on the equilibrium simulation path – deteriorates
substantially, especially at longer horizons, when log std(ε) is not used.
Table 25: Economic Implications of log std(ε) Being in the Krusell-Smith Rules
correlation between Y BL-Model w log std(ε) 0.53and std(ij,t) BL-Model w/o log std(ε) 0.28
Data 0.45
Y C I Nvolatility BL-Model w log std(ε) 2.19 0.82 10.22 1.52
BL-Model w/o log std(ε) 2.16 1.03 10.23 1.56Data 2.30 1.79 4.37 1.80
persistence BL-Model w log std(ε) 0.29 0.55 0.22 0.20BL-Model w/o log std(ε) 0.22 0.65 0.08 0.06Data 0.48 0.67 0.42 0.61
correlation with Y BL-Model w log std(ε) 1.00 0.87 0.98 0.96BL-Model w/o log std(ε) 1.00 0.74 0.93 0.89Data 1.00 0.66 0.83 0.68
Notes: The table displays – in the upper panel – the correlation of the dispersion of the investment rates
with the cycle for the model under the baseline calibration (‘BL-Model’), where log std(ε) is included in
the Krusell-Smith rules, a version of the baseline model, where log std(ε) is not included in the Krusell-
Smith rules, as well as from the USTAN data. In the lower panel the table does the same comparison for the
percent standard deviations (volatility), autocorrelation (persistence), and correlation with aggregate output
of HP(100)-filtered log aggregate output (Y), consumption (C), investment (I), and employment (N).
49
Table 26: Quality of KS Rules - R2 and the Standard Error
Rule for k Rule for p
KS rules include log std(ε)
s1 s2 s3 s4 s5 s1 s2 s3 s4 s5
z1 - - - 0.9999 1.0000 - - - 0.9996 1.0000z2 - 0.9997 0.9996 0.9997 1.0000 - 0.9999 0.9991 0.9994 0.9998z3 0.9997 0.9993 0.9997 0.9997 1.0000 0.9997 0.9989 0.9994 0.9995 1.0000z4 0.9995 0.9996 0.9997 0.9997 - 0.9994 0.9994 0.9996 0.9999 -z5 0.9997 0.9997 - - - 0.9995 0.9997 - - -
Results for the regression w/o all interaction effects
R2 = 0.9998 S.E. = 3.14 ∗ 10−4 R2 = 0.9997 S.E. = 2.20 ∗ 10−4
KS rules do not include log std(ε)
s1 s2 s3 s4 s5 s1 s2 s3 s4 s5
z1 - - - 0.9586 0.9769 - - - 0.9984 0.9989z2 - 0.9420 0.8786 0.8782 0.8890 - 0.9992 0.9959 0.9960 0.9958z3 0.8256 0.8241 0.8729 0.8763 0.9029 0.9952 0.9935 0.9956 0.9956 0.9970z4 0.8398 0.8745 0.8878 0.9592 - 0.9940 0.9952 0.9959 0.9985 -z5 0.9376 0.7452 - - - 0.9972 0.9886 - - -
Results for the regression w/o all interaction effects
R2 = 0.9062 S.E. = 8.00 ∗ 10−3 R2 = 0.9976 S.E. = 8.00 ∗ 10−4
Notes: All simulation results reported in the paper refer to the setup in the upper panel.
50
Table 27: Quality of Forecasting Rules at Various Horizons
Forecast Horizon tRule for . . . Evaluation Criterion t = 1 t = 5 t = 10 t = 100
k Mean Squared % Dev. forecast - actual 0.0003 0.0007 0.008 0.0008k Max. Abs. % Dev. forecast - actual 0.0014 0.0034 0.0036 0.0037k Correlation forecast - actual 0.9999 0.9995 0.9994 0.9994
p Mean Squared % Dev. forecast - actual 0.0002 0.0003 0.0003 0.0003p Max. Abs. % Dev. forecast - actual 0.0007 0.00010 0.0011 0.0011p Correlation forecast - actual 0.9999 0.9998 0.9997 0.9997
Table 28: Quality of Forecasting Rules at Various Horizons - KS rules donot include log std(ε)
Forecast Horizon tRule for . . . Evaluation Criterion t = 1 t = 5 t = 10 t = 100
k Mean Squared % Dev. forecast - actual 0.0080 0.0220 0.0270 0.0296k Max. Abs. % Dev. forecast - actual 0.0273 0.0832 0.0966 0.0943k Correlation forecast - actual 0.9525 0.5995 0.3689 0.2404
p Mean Squared % Dev. forecast - actual 0.0008 0.0087 0.0116 0.0130p Max. Abs. % Dev. forecast - actual 0.0027 0.0309 0.0402 0.0394p Correlation forecast - actual 0.9989 0.8468 0.7201 0.6471
Notes: All simulation results reported in the paper refer to the setup in Table 27.
51